Fagerfjall Final PDF

MASTER’S THESIS
Optimizing wind farm layout

– more bang for the buck using mixed
integer linear programming
Patrik Fagerfjäll
Department of Mathematical Sciences

CHALMERS UNIVERSITY OF TECHNOLOGY
GOTHENBURG UNIVERSITY
Göteborg, Sweden 2010
Thesis for the Degree of Master of Science
Optimizing wind farm layout

– more bang for the buck using mixed
integer linear programming
Patrik Fagerfjäll
Department of Mathematical Sciences

Chalmers University of Technology and Gothenburg University
SE-412 96 Göteborg, Sweden
Göteborg, June 2010
Mathematical Sciences
Göteborg 2010
Abstract
Designing a wind farm layout is as of today mainly performed manually.

Many different factors affecting the revenue of a wind farm and a large
amount of information must be considered, making a close to optimum lay-
out a hard problem to solve manually. In this thesis, mixed integer linear
programming models capable of optimizing many common layout problems
are developed.
We have formulated two optimization models, the Wind Farm Production

Optimization model and the Wind Farm Infrastructure Optimization model.
In the Production model, wind turbines are positioned with respect to a
minimum separation distance. Furthermore, production losses due to wake
effects between wind turbines are accounted for. The model also superpo-
sitions sound pressures in nearby areas for different wind directions and is
also capable of maximizing profit instead of maximizing the production. The
Infrastructure model connects the wind turbines by roads and cables, where
the latter can include choosing cable dimensions and calculating cable power
losses relating to the current. The model can also optimize the positioning
of the transformer station.
The two models are capable of optimizing the problems stated at a reasonable
computation time. The Production model was compared to the commercial
WindPRO 2.6 Optimize module in verification examples housing 20 to 30
wind turbines. For the problem of maximizing the total production in the
given geographic areas, the Production model of this thesis managed to find
locations for many more wind turbines than Optimize, yielding about 40%
higher total production. When restricted to allow only as many wind tur-
bines as the software Optimize was able to place, the Production model still
performed equal or better.
Our tests show that the field of mixed integer linear programming and the
mathematical models of this thesis possess a great potential to aid in the
process of increasing the return of wind farms. These tools are versatile and
can be adopted to many different scenarios. They need further development
to be able to handle really large wind farm projects but are still, as of today,
capable of delivering more bang for the buck in wind farm layout design.
i
Sammanfattning
The following is an extensive Swedish abstract.
Målet med detta examensarbete är att, givet beräkningsmetoder för produk-

tion, vakeffekter, ljudutbredning och kostnader, skapa matematiska modeller
för att optimera total produktion, kostnader eller vinst vid designen av en
vindkraftparks layout. Arbetet utvärderar inte olika fysikaliska modeller el-
ler andra optimeringsmetoder, men jämför en av de skapade modellerna med
motsvarigheten i den kommersiella programvaran WindPRO 2.6 Optimize
genom exempelproblem, dock ej genom någon kvantitativ studie.
Layouten hos en vindkraftpark designas idag ofta för hand. Som underlag
har man ett projektområde och utifrån en karta över terräng eller beräknad
energitäthet skall man placera ut vindturbiner för att nå en hög total pro-
duktion. En liten ändring av en vindturbins medelvind ger stora skillnader
i effektuttag, och en effektiv inbördes placering är därmed viktigt för att få
bra avkastning från parken. Lokala variationer i vindstyrka och vindfrekvens i
olika riktningar för olika delar av ett parkområde medför mycket information
att hålla reda på. Lägg till detta hur vakeffekterna mellan två vindturbiner
påverkar dessas produktion negativt samt hur ljudnivåer från flera vindtur-
biner samverkar till att öka minsta tillåtna avståndet till bebyggelse, och
processen blir väldigt detaljrik. Utöver dessa produktionsaspekter innefattar
designen också problem som att välja rätt dimension av kablar samt att dra
dessa och interna vägar på billigast möjliga sätt.
Två modeller skapas, av typen linjära heltalsprogram, där den ena repre-
senterar problemet att maximera produktion och den andra problemet att
minimera kostnader för kablar, vägnät och placering av transformatorsta-
tion. Båda modellerna kan beskrivas med hjälp av grafer, vilka består av en
mängd geografiskt utspridda noder samt bågar som sammanbinder dessa. No-
derna representerar möjliga placeringar av vindturbiner, där en nod i vilken
en vindturbin placerats kallas aktiv, och bågarna symboliserar vakeffekter,
kablar eller vägar mellan noder. En aktiv båge i Infrastruktursmodellen kan
exempelvis motsvara en kabeldragning mellan två vindturbiner.
ii
I Produktionsmodellen implementeras ett minsta tillåtna avstånd mellan två
vindturbiner som ett tätpackningsproblem av cirklar vars radier är halva det-
ta minsta avstånd. Två placerade vindturbiner medför produktionsförluster i
form av vakeffekter vilket representeras av en aktiv båge mellan motsvaran-
de noder i grafen. Den kombinerade effekten av vakeffekter från flera källor
beräknas genom linjär överlagring. Då andra kombinationsmetoder tidigare
har visat sig stämma bättre överens med verkligheten, exempelvis att sum-
mera kvadraterna av hastighetsförlusterna, skulle detta kunna vålla problem
för Produktionsmodellen, och en metodik för att i dessa modeller approxi-
mera andra samband än linjär överlagring presenteras. Vidare kompletteras
modellen med bivillkor som begränsar den kombinerade ljudnivån i en viss
punkt från alla vindturbiner och för olika vindriktningar. Detta innebär att
det minsta avståndet från en vindturbin till en närliggande bostad beror av
antalet vindturbiner, i vilka riktningar från bostaden de är placerade samt
hur vindkarakteristiken för dessa olika riktningar ser ut. Denna modellering
medför större möjligheter än om man endast anger ett minsta avstånd till
bebyggelse, vilket görs i Optimize.
I Infrastrukturmodellen består bågarna i grafen av möjliga vägval för kabel-

och vägdragning. Här skapas en trädstruktur av effektflöden från vindturbi-
nerna till den gemensamma transformatorstationen. Bågarna tillåts även att
ansluta till icke aktiva noder för att möjliggöra förgreningar utanför vindtur-
binernas positioner. Slutna slingor i trädet, vilket är ett problem som måste
hanteras vid trädoptimering, undviks då detta effektflöde endast kan skapas i
vindturbinerna och allt flöde måste nå transformatorn. Från transformatorn
skickas effektflödet i den dyrare exportkabeln till valfri anslutningspunkt. Mo-
dellen kan behandla kabeltyper med olika kapacitet samt modellera dubbla
kablar och de effektförluster ström genom en kabel ger upphov till. De olika
typerna modelleras genom att varje båge i grafen kan representeras av flera
matematiska variabler med olika övre gräns för mängden effektflöde och olika
kostnader. Utöver kostnaderna för kablar och vägar ingår i Infrastrukturmo-
dellen terräng- och platsberoende kostnader för fundament till transforma-
torstation samt för kabelförgreningar i de fall dessa sker i kopplingsstationer
istället för vid befintlig vindturbin.
De båda modellerna kan genom att prissätta produktionen kombineras till

en gemensam modell med målet att maximera avkastningen, vilket typiskt
leder till att enskilda vindturbiner inte byggs för långt ifrån de resterande
eller på en plats dit vägbygge eller kabeldragning blir för dyrt.
Utöver de implementerade modellerna föreslås tillägg för att öka hastighe-
iii
ten med vilken Produktionsmodellen kan lösa problem med hög upplösning.
Denna höga upplösning har mindre betydelse för storleken på vakeffekterna
än för möjligheten till precis och tät placering av vindturbinerna. För att
antalet bågvariabler i grafen därför inte skall växa som kvadraten på antalet
noder, föreslås en uppdelning av ytan i sektorer. Vindturbiner kan placeras
noggrant över det högupplösta rutnätet, medan vakeffekter samt eventuellt
kablar och vägar approximeras med bågar mellan de sektorer som innehåller
vindturbiner. Antalet bågar motsvarar då antalet sektorer i kvadrat obero-
ende av upplösningen. Storleken på dessa sektorer väljs så att det minsta
tillåtna avståndet mellan vindturbiner förbjuder fler än en vindturbin per
sektor samt så att vakeffekterna mellan sektorer förblir en god approxima-
tion av vakeffekterna mellan de riktiga positionerna för vindturbinerna.
De utvecklade modellerna visar sig fungera väl för sina respektive ändamål
och kan användas i processen att skapa en layout för en vindkraftpark. Pro-
duktionsmodellen kan optimera vindturbinernas placeringar med avseende på
antingen produktion eller vinst, och kan därmed bidra till en högre total pro-
duktion och större intäkter. Infrastrukturmodellen hittar prisvärda lösningar
på kabeldragningar med val av kabeltyp samt vägnät, vilket kan bidra till
att minska kostnaderna för dessa. Då båda modellerna kan hantera mycket
av den ingående problematiken kan också en större del av layoutprocessen
automatiseras och därmed tiden för det manuella arbetet minskas. Kombi-
nationen av de båda modellerna till en enda optimeringsmodell är intressant
ifall kostnaderna för kabel- och vägnät är av betydande storlek relativt vins-
ten från vindturbinernas produktion, då förbättringar hos dessa annars en-
dast leder till obetydliga förbättringar i totala vinsten och vindturbinernas
position sannolikt inte påverkas. Relationen mellan produktionsvinsten och
kostnaden för kabel- och vägnät beror på elpris samt vilken önskad återbetal-
ningstid för parken som används. Den kombinerade modellen blir dock stor
och tar lång tid att lösa och vissa föreslagna genvägar för att minska antalet
noder i infrastrukturmodellen inte kan användas.
För verifiering av Produktionsmodellen i ett riktigt projektområde jämför-

des lösningarna funna efter en timme för två upplösningar med lösningarna
funna genom den kommersiella optimeringsprogramvaran WindPRO 2.6 Op-
timize. Som facit på de placerade vindturbinernas produktion genomfördes
produktionsberäkningar för de framtagna layouterna med hjälp av verktyget
WindPRO 2.6 PARK, vilket är ett verktyg för att beräkna total produk-
tion för en viss layout. Båda metoderna fick samma indata, använde samma
modeller för produktion- och vakberäkningar och optimerade med målet att
maximera total produktion. Kombinationen av multipla vakeffekter hanteras
iv
dock på olika sätt: Produktionsmodellen som överlagrar enskilda vakeffekter
överskattar den kombinerade förlusten från två vakar jämfört med Optimize
och PARK som båda kombinerar vakar som summan av kvadraterna av has-
tighetsförlusterna. Då Optimize inte kan hantera överlagring av ljudintensitet
från flera vindturbiner, utan endast använder ett minsta tillåtna avstånd från
var och en av dessa, skalades Produktionsmodellen ner till att inte inklude-
ra denna del. Jämförelsen visade att Produktionsmodellen finner en högre
total produktion än Optimize. Detta då den kunde placera ut många fler
vindturbiner än vad Optimize lyckades med. I fallet med en upplösning om
100m lyckades Optimize placera ut endast 18 vindturbiner medan motsva-
rande antal för Produktionsmodellen var 28 stycken, vilket innebar en 43%
högre total produktion. För upplösningen 50m var motsvarande antal vind-
turbiner 20 respektive 30 stycken, vilket gav en 38% högre total produktion.
De av Produktionsmodellen utplacerade antal vindturbiner motsvarar ock-
så de maximala antalen för respektive upplösning. Om antalet vindturbiner
begränsades i Produktionsmodellen till de 18 respektive 20 stycken som Op-
timize lyckades placera ut gav denna en total produktion som var ungefär en
procent högre än Optimize i problemet med upplösningen 100m och ungefär
lika i problemet med upplösning 50m. I det senare problemet begränsades
Produktionsmodellen av programvarans tilldelade minne till att inte kunna
inkludera vakeffekter på längre avstånd än 900m. Trots att Optimize och facit
PARK kombinerar multipla vakeffekter annorlunda än Produktionsmodellen
gav denna alltså lösningar med lika bra eller bättre total produktion.
De utvecklade modellerna kan hitta värdefulla förbättringar i många av de

delmoment som ingår i en designprocess, och kan dessutom spara in på den tid
det skulle tagit att skapa dessa för hand. De kan hantera projekt på land och
offshore, med fri positionering eller bunden av visuella aspekter eller specifika
möjliga positioner. Produktionsmodellen kan hantera avancerade ljudutbred-
ningsmodeller som beror av vindriktningar. Infrastrukturmodellen har också
kompletterats med tillägg som möjliggör att optimera sammankoppling av
vindkraftparker istället för enskilda vindturbiner, vilket möjliggör användan-
de av denna modell i ett vidare sammanhang. Programmen behöver komplet-
teras och testas, dels med föreslagna förenklingar så som sektoruppdelning
av vakvariabler, dels med förfinade optimeringsparametrar, såsom bättre val
av branching strategy, för att klara av att lösa riktigt storskaliga problem
på rimlig beräkningstid. För parker på ett antal kvadratkilometer med några
tiotals vindkraftverk går problemen dock redan nu att optimera till tillräcklig
nivå på någon timme på en persondator. Linjär heltalsprogrammering visar
sig vara en fungerande teknik för att optimera dessa layouter och ytterli-
gare utveckling bör utföras inom området. Modellerna kan redan idag, och
v
kommer än bättre i framtiden kunna, skapa layouter som medför en bättre
avkastning från vindkraftparker.
vi
Contents
Contents vii
1 Introduction 1
1.1 Structure of the thesis . . . . . . . . . . . . . . . . . . . . . . 2
1.2 The need for wind farm layout optimization . . . . . . . . . . 2
1.3 From a mathematical point of view . . . . . . . . . . . . . . . 3
2 Wind farm layout 5
2.1 Maximizing production in a given area . . . . . . . . . . . . . 6
2.1.1 Project site area . . . . . . . . . . . . . . . . . . . . . 6
2.1.2 Wind distribution . . . . . . . . . . . . . . . . . . . . . 6
2.1.3 Minimum distance between wind turbines . . . . . . . 7
2.1.4 Production loss between wind turbines . . . . . . . . . 9
2.1.5 Sound levels in surrounding areas . . . . . . . . . . . . 10
2.1.6 Choice of wind turbine . . . . . . . . . . . . . . . . . . 10
2.2 Building infrastructure and reducing costs . . . . . . . . . . . 10
vii
2.2.1 Foundations . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2 Cables . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.3 Roads . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2.4 Transformer station . . . . . . . . . . . . . . . . . . . . 13
3 About integer linear programming - the mathematical mod-

eling tool 15
3.1 Linear programming . . . . . . . . . . . . . . . . . . . . . . . 16
3.2 Integer linear programming . . . . . . . . . . . . . . . . . . . 19
3.3 Graphs and trees . . . . . . . . . . . . . . . . . . . . . . . . . 22
4 Formulating the optimization model 27
4.1 What to formulate . . . . . . . . . . . . . . . . . . . . . . . . 28
4.2 Placing wind turbines and maximizing production . . . . . . . 28
4.2.1 Possible production at different locations . . . . . . . . 28
4.2.2 Minimum distance between wind turbines . . . . . . . 29
4.2.3 Packing density in a discrete grid . . . . . . . . . . . . 33
4.2.4 Production loss between wind turbines . . . . . . . . . 33
4.2.5 Maximum sound levels . . . . . . . . . . . . . . . . . . 35
4.2.6 The complete Production model . . . . . . . . . . . . . 36
4.2.7 Additional formulations . . . . . . . . . . . . . . . . . 37
4.3 Infrastructure formulation . . . . . . . . . . . . . . . . . . . . 42
viii
4.3.1 Foundation costs . . . . . . . . . . . . . . . . . . . . . 42
4.3.2 Formulating a Steiner tree . . . . . . . . . . . . . . . . 42
4.3.3 Minimizing the costs of roads . . . . . . . . . . . . . . 47
4.3.4 Cables . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
4.3.5 Transformer station placement . . . . . . . . . . . . . . 53
4.3.6 The complete Infrastructure model . . . . . . . . . . . 53
4.3.7 Additional formulations . . . . . . . . . . . . . . . . . 55
4.4 Combined revenue formulation . . . . . . . . . . . . . . . . . . 60
5 Tests and results 63
5.1 Verification example and comparison to commercial software . 65
5.1.1 Optimization of an actual project . . . . . . . . . . . . 65
5.1.2 Comparison to commercial optimization software . . . 69
5.2 Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
6 Conclusion and recommendations 85
6.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
6.2 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
6.3 Recommendation for future work . . . . . . . . . . . . . . . . 91
Bibliography 93
ix
x
Acknowledgements
This master’s thesis was performed in collaboration with Vattenfall Power

Consultant at their facilities. Firstly I want to thank the two people at this
company who initially trusted in me and granted me the possibility to un-
dertake an optimization work of my choice – my supervisor Lasse Johansson
and my boss Mattias Törnkvist. I am also grateful for them allowing me to
take part in all various types of social and professional events.
I wish to thank all of my co-workers at the Wind unit of Vattenfall Power

Consultant, who have supported my work with ideas and with answers to
numerous questions, as well as provided me with a very likable working en-
vironment. Out of these, I especially wish to extend my gratitude towards
my two supervisors, Lasse Johansson and Anna Larsson, for their aid and
expertise.
For the mathematical input and discussions, as well for the extensive reading
and improvement of this theses, I would like to thank my supervisor and
examiner, Docent Ann-Brith Strömberg at the department of Mathematical
Sciences at Chalmers University of Technology.
Göteborg, May 2010
Patrik Fagerfjäll
xi
CHAPTER 1. INTRODUCTION
Chapter 1
Introduction
1
1.1 Structure of the thesis
This thesis describes the work of developing models for addressing the opti-
mization of production and cost of the wind farm layout optimization prob-
lem. It relies on existing methods and models for calculating wind resources
in a geographic area, production, wakes, sound propagation and so on, and
is a method for combining these models to optimize the locations of wind
turbines and the road and cable networks. The thesis does not evaluate
physical models or other means of optimization. It will through verification
examples compare one of the models developed to the corresponding module
in the commercial software WindPRO, however, no extensive quantitative
study is undertaken. The aim of the thesis is to develop models and to con-
clude whether it is possible to increase production and profits, rather than
to indicate by how much these may be increased.
The thesis starts by listing many important subproblems of the wind farm
layout design problem. It then gives an introduction to the modeling tool
used in this thesis, mixed integer linear programming. Following these initial
chapters, the two models - the Wind Farm Layout Production Optimization
model and the Wind Farm Layout Infrastructure Optimization model - ca-
pable of solving the problems stated are built, some performance results are
listed and the Production model is verified and compared with commercial
software in a real life example. The application of the models in a real layout
processes is the main focus of the Discussion section.
1.2 The need for wind farm layout optimization
As of today much of the wind farm layout process is done manually, partly
since there does not seem to exist any good optimization tools (one commer-
cial optimization tool will be presented in this thesis). Care should be taken
so that the iterative work flow of a design process is not restricted too much.
However, due to the enormous amount of information that must be handled
when searching for the best wind turbine locations, there is a lot of room for
computer aided optimization. Although the uncertainties in the models used
for modeling, for example, wind flows and the spread of a wake, need not be
small [22], this is the information that financing decisions are currently based
upon. Thus, increasing the modeled total profit is of great value. Since this
2
profit is highly dependent on the power production, and thus the positioning
of the wind turbines, an effort creating tools that yield a higher production,
even if merely a percent higher, would be well worth.
1.3 From a mathematical point of view
The wind farm layout optimization problem consists of several theoretically

interesting subproblems. Circles are to be packed tightly in an area and
nodes in a graph are to be connected by a spanning tree with the possibility
but not the obligation of using additional nodes to shorten the total length of
the tree, which transforms the spanning tree problem into a so called Steiner
tree problem. Optimization in wind farm layout design is a young field of
application and while heuristic or genetic algorithms seem to have been the
first solution methods chosen [7, 19, 17, 12, 23] it is important to investigate
the performance of an integer linear programming approach.
3
4
CHAPTER 2. WIND FARM LAYOUT
Chapter 2
Wind farm layout
5
Designing a wind farm layout is a process involving many branches of exper-

tise, including wind analysis, construction, power grid dimensioning, envi-
ronmental impact analysis and in some sense politics. This chapter will list
a few of the considerations important in the layout process, which will all
be incorporated in the model created. They are not exclusive, but represent
most of the important subproblems of wind farm layout design.
2.1 Maximizing production in a given area
The first and major part of the wind farm layout optimization problem is to
find the optimum wind turbine locations. This section will define important
factors affecting permitted locations and wind farm production.
2.1.1 Project site area
Many aspects are taken into account when finding and deciding on project
site area. These aspects include finding a windy and accessible piece of land,
not covered by, for example, wetlands or any type of nature reserve. More-
over, there could be considerations regarding the site area posing as a habitat
to endangered species or being the nesting grounds for birds. Restrictions re-
garding minimum distance from, for example, existing roads could also apply
and should be accounted for when defining the project area. In this thesis it
is assumed this area is defined with restricted areas removed.
2.1.2 Wind distribution
The power in a wind is proportional to the cubic wind speed, v 3 , where v 2

comes from the kinetic energy of the air mass and v from the air mass flow
(Figure 2.1). Thus, a small change in wind speed can have great impact
on produced electrical output. Expressed relative to the wind speed, a one
percent increase of the wind speed v yields a (1.01)3 − 1 ≈ 3% gain in v 3 and,
thus, production. A good site location and proper wind turbine placement
within the wind farm are therefore of importance.
6
18000 60
16000
50
Shift in cubic wind speed [%]

14000
Cubic wind speed
12000 40
10000
30
8000
6000 20
4000
10
2000
0 0
0 5 10 15 20 25 30 0 2 4 6 8 10 12 14 16
Wind speed Shift in wind speed [%]
Figure 2.1: Power in wind proportional the cube of the wind

speed.The right graph describes the percentage increase in cubic wind speed
per percentage increase in wind speed, (∆v, (v + ∆v)3 /v 3 ). The fitted linear
line has an incline of about 3.5.
Wind speed distribution and directions are calculated for all of the site area.
This calculation is based on wind measurements made over a longer time,
preferably from a hub height mast at the project site, which are corrected for
deviance from yearly averages. Photos showing two means of measuring wind
characteristics, a mast and a sodar, are shown in Figure 2.2. Softwares used
in the modeling could be based on computational fluid dynamics (CFD) or
some terrain flow model, and incorporate, for example, elevation, roughness
of the ground surface for a large area surrounding the site [18, 8]. The
output of this software and the input to the thesis models will be Weibull
distribution parameters in a number of angular sectors for all points in a
grid covering the project area and surroundings. The Weibull distribution
has been shown to represent wind speed frequencies well and is commonly
used [24]. A Weibull distribution along with an example of a power curve and
the resulting production distribution is shown in Figure 2.3. An illustrative
representation of the Weibull parameters and wind direction frequency for a
number of sectors is the wind rose, seen in Figure 2.4.
2.1.3 Minimum distance between wind turbines
Wind turbines cannot be placed too close to each other. The most obvious
distance is two rotor radii apart to prevent the rotor blades from collid-
ing. There are however usually other requirements or demands on minimum
7
(a) Mast (b) Sodar
Figure 2.2: Means for measuring wind speed. Picture (a) shows an 80m
high mast fitted with anemometers whereas in picture (b) a sodar, measur-
ing wind speed at different heights by observing mainly the doppler shift
of returning aucustic impulses, is seen. Both pictures are taken in Sourva,
Sweden.
14 2500 120
Probability Density Function
Relative power output [kW]
12 100
2000
Power output [kW]
10
80
8 1500
60
6 1000
40
4
500 20
2
0 0 0
0 10 20 30 0 10 20 30 0 10 20 30
Wind speed [m/s] Wind speed [m/s] Wind speed [m/s]
Figure 2.3: Weibull distribution, a power curve example and the re-
sulting production distribution. The latter is the product of the Weibull
distribution and the power curve at each wind speed.
8
Frequency [%] A parameter k parameter

1 1 1
15 8 2,5
12 2 12 2 12 2
6 2
10
1,5
11 3 11 4 3 11 3
1
5
2 0,5
10 0 10 4 0 10 4 0 4
9 5 9 5 9 5
8 6 8 6 8 6
7 7 7
Figure 2.4: Wind roses, showing frequency and Weibull A and k parameters
for twelve 30 degree sectors, where north is denoted by 1.
separation distance to be met made from manufacturers, partly to prevent

excessive tear on the construction due to turbulent winds [15, 24]. A typical
minimum distance could be around four or five rotor diameters but would
sometimes be set higher when manually placing the wind turbines due to
production losses between them, further discussed in the next section.
2.1.4 Production loss between wind turbines
Wind passing through a wind turbine will decrease in speed, not to recover
fully for quite some distance. Thus, two wind turbines standing not too
far apart will have a negative effect on each others production, referred to
as wake effects or array losses. This loss of wind speed could be calculated
through a number of models and the optimal positioning of wind turbines will
depend on this effect. These wake models are often based on single turbine
wakes, and the method for combining several wakes varies. One common
way is to sum the kinetic energy deficits, that is the sum of the square of the
velocity deficits. This is said to in some cases have a better resemblance with
reality than just linear superposition, although not generally accepted why
[14, 22, 9]. This thesis will however, due to the linear programming approach,
make use of linear superposition and the differences between these methods
will be addressed later on. For the purpose of manually positioning wind
9
turbines a distance between two turbines of 5 rotor diameters perpendicular

to, and 7 rotor diameters parallel to the prominent direction of wind is often
proposed [24].
2.1.5 Sound levels in surrounding areas
There are regulations stating maximum sound levels originating from wind
turbines at nearby houses or other sensitive areas. Through the use of models
calculating sound levels around a given source, this relates to a minimum
possible distance between wind turbines and these areas. The regulations
could be formulated in a number of ways, as, for example, "must not exceed
an equivalent sound level of 40 dB(A) at nearby houses, and superposition
of sound pressure from multiple wind turbines is made as if the sensitive
area is downwind of all wind turbines simultaneously", partly taken from
the Swedish Naturvårdsverket [1]. The minimum distance will however vary
depending on the number of wind turbines positioned. Two wind turbines of
the same sort in about the same spot will yield double the sound pressure,
equivalent to a 3dB sound level rise, in a nearby point. This effectively
increases the minimum distance. Sound pressure and sound level relates as
SL = 10 log10 (SP).
2.1.6 Choice of wind turbine
The choice of wind turbines will affect the production in a location, the
magnitude of the wake effects and the sound levels. In many wind turbines
there is also an option to decrease the latter at the cost of also decreasing
power production.
2.2 Building infrastructure and reducing costs
It is not only the available amount of wind energy in each point, and thus the
possible power output, that will decide the optimal wind turbine locations
and return on investment. The cost of constructing a wind farm could vary
depending on the amount of work and material needed for foundations as
10
well as roads and cables connecting the wind turbines, transformer station
and the external power grid. The problem of optimizing this is in contrast to
the above aspects of maximizing production a problem of minimizing costs.
Relative to the cost of wind turbines the costs of roads and cables are rather
small and it is common these costs will not be considered until after the wind
turbine locations have been chosen [24]. If locations are already decided, a
model optimizing the costs could be formulated and used straight away. If,
however, these costs are to be considered when finding the optimal wind
turbine locations, other approaches are needed since the two problems are
optimizing different quantities. One could optimize over both production
and costs by letting one of them be fixed at a certain level, and do this over
and over again using different levels to create a type of trade-off curve, in
economics often referred to as a Pareto efficient frontier. Another approach
is to bridge the two problems by defining a revenue on electricity production,
thus optimizing revenue minus costs over a given time period. This latter
approach will be implemented for the purpose of testing problem complexity.
2.2.1 Foundations
Wind turbines should not be built where the terrain, for example, is too
steep, too rough or consists of too unstable ground for it to be economically
beneficial or even possible to use as a location. To deal with this, the cost of
foundation must be priced correctly in different areas. One should also add
the cost of each wind turbine, though this cost will generally not vary due
to the choice of location. If using multiple types of wind turbines, the costs
of these and possibly foundation costs could vary depending on which type
is chosen. Figure 2.5 shows pictures from the construction of a small wind
farm.
2.2.2 Cables
Wind turbines need to be connected to a transformer station and the ex-

isting external grid, and for this interconnecting cables buried underground
are generally used. Since the act of transporting and burying a cable re-
quires a road to be constructed along its path, roads and cables are usually
constructed taking the same paths.
11
(a) Nacelle (b) Construction
Figure 2.5: The construction of a wind turbine, with hub height 105m.
Both pictures are taken in Simmatorp, Sweden.
Interconnecting cables of different dimensions can allow for different current

levels and thus different number of connected wind turbines. Several cables
can also be placed along side one another requiring only one road to be built
and affecting the amount of power losses. The model should thus be able to
incorporate different types of cable on each path. The cost of a cable will
depend on both the work needed for constructing the road and the burying
as well as the type of cable selected. For connecting the transformer station
to the external grid, an export cable with high capacity and cost is used. In
addition to the construction costs there is the cost of power losses which is
proportional to the length of a cable and the square of the current.
Several cables originating from wind turbines could be combined into high
capacity cables, as in the branches of a tree, with additional costs for con-
nector stations if this connection is not performed at another wind turbine.
If defining a power flow from the wind turbines to the transformer station,
flow can be combined but not split.
2.2.3 Roads
Roads connecting the wind turbines and the outside world are needed for
transporting the wind turbines to their locations and performing mainte-
12
nance, where the former is the need defining what roads can be built or not.
Wind turbine transports are heavy, with the generator as the heaviest part,
or really long, as when transporting the blades (as seen in Figure 2.6). This
calls for robust, low-incline roads, which, for example, could result in it not
being possible to build on top of a hill even though wind energy levels might
be high and foundation costs low.
Figure 2.6: Wind turbine blade transport. Each blade is 45m long. The
picture is taken in Simmatorp, Sweden.
2.2.4 Transformer station
A large wind farm would usually connect to the external power grid through
a transformer station which, just as in placing a wind turbine, will result
in foundation as well as road and cable costs along with the cost of the
transformer station itself. Its optimal placement is dependent on the cost of
the high voltage export cable needed for connecting to the external power
grid compared to the cables needed for connecting the wind turbines. The
construction cost of the transformer station can also vary depending on its
location within the farm. Power losses in the transformer will depend on the
current, and thus the number of wind turbines, but not on the location of
the wind turbines within the farm. It is possible to allow for different sizes
of transformer stations if the number of wind turbines to be placed is not
decided prior to optimization.
13
14
CHAPTER 3. ABOUT INTEGER LINEAR PROGRAMMING - THE
MATHEMATICAL MODELING TOOL
Chapter 3
About integer linear

programming - the mathematical
modeling tool
15
This section gives a short introduction to the field of mixed integer linear
programming used in developing the models of the thesis. Examples and
basic theory of linear programs, integer linear programs and the definition
of graphs and trees are presented, along with a final section defining the
notations used in this thesis.
3.1 Linear programming
The word linear in linear programming (LP) refers to the variables of a prob-
lem being combined in a linear way, that is adding or subtracting variables
but not, for example, multiplying them or using them as exponents. Most
physical problems, or other types of problems for that matter, cannot be
expressed in a linear way, but those who can can be solved in more efficient
ways. The concept of linear optimization will here be explained by the use
of an example.
Let us say we have discovered that the citizens of New York have great
cravings for Swedish Knäckebröd and Messmör, a form of crispbread and
soft whey butter. We see a great business opportunity and decide to bring a
cart and a bag full of these goods across the Atlantic to pursue a career as
Knäckebröd and Messmör vendor. On the flight to America the cart will take
up all of our checked baggage allowance, leaving us with just the carry-on
baggage for transporting our goods. If the economic and physical properties
of Knäckebröd and Messmör are according to Table 3.1 and the carry-on
baggage is restricted to 8 kg and 60 litres respectively, what quantities of
the two products should we bring to maximize our profit in the land of
opportunities? For the moment, fractions of packages are allowed to be
bought and sold and we are able to pack the goods real tight in our baggage.
[per package] Knäckebröd Messmör

Cost [SEK] 20 35
Revenue [SEK] 40 60
Profit [SEK] 20 25
Weight [kg] 0.5 0.8
Volume [litre] 7.5 0.6
Table 3.1: The properties per package of Knäckebröd and Messmör
16
The total potential profit, which is what we want to maximize, can be ex-
pressed as the linear function of packages of Knäckebröd, K, and Messmör,
M , called the objective function, to
maximize 20K + 25M.
Since this objective function is ever increasing, the constraints of maximum

weight and volume of our combined goods have to be added. This is done
by introducing two more linear functions, and the linear program reads
maximize 20K + 25M,

subject to 0.5K + 0.8M ≤ 8,
7.5K + 0.6M ≤ 60.
Now, since we should not be able to buy or export a negative number of

goods, neither of the variables K and M are allowed to obtain a value less
than zero (another effect of allowing negative variable values would be the
possibility to sell, for example, Messmör before entering the aircraft, yielding
a negative impact on the total weight and volume, thus, making it possible
export more than 8 kg and 60 litre of Knäckebröd). The added constraints
now yields the linear program to
maximize 20K + 25M,

(3.1)
7.5K + 0.6M ≤ 60,
K, M ≥ 0.
The possible combinations of values, the set of feasible solutions, of K and

M not violating the constraints can now be illustrated by the bounded area
of Figure 3.1. All other combination of variables are referred to as infeasible.
Both the variables need to be non-negative and are restricted by the two lines
representing the weight and volume constraints. This type of area, bounded
by straight lines, is a called a polyhedron and is in this instance convex. Con-
vexity basically means, when referring to an object in an Euclidean space,
that there are no holes or dents in the object - the straight line segment
joining any pair of points within the object, lies entirely within the object.
Mathematically speaking, a set U is convex if for any u1 , ..., un ∈ U , and any
17
λ1 , ..., λn ≥ 0, such that nk=1 λk = 1, the vector nk=1 λk uk ∈ U . Linear

P P
functions on such convex polyhedrons, or in several dimensions convex poly-
hedrons, have the property of attaining their maximum and minimum values
in an extreme point of the polygon or along one of its sides (that is, along the
feasible region of a constraint line) [16]. In Figure 3.1 these possible extreme
points are marked by circles, and the act of optimizing the problem 3.1 is
now a matter of locating the right extreme point. In this two-dimensional
problem, this can be seen by just observing the figure while knowing which di-
rection is the gradient of increasing profit. For a problem consisting of many
more variables, optimization through such a visual approach is however not
possible and mathematical methods are required.
18
16
14
weight = 8kg
12
knäckebröd, k
10
8 volume = 60l
6
4
2
0
0 2 4 6 8 10 12
messmör, m
Figure 3.1: A convex polyhedron of feasible solutions. Due to the

linearity of the objective and constraints, and the set of feasible solutions
being a convex polyhedron, the optimum will be in one of the extreme points
or along a constraint line. The gradient of increasing profit is illustrated by
an arrow.
The optimal solution to the problem of exporting Knäckebröd and Messmör

is shown in Figure 3.2. In this particular example, at this intersection of
constraint lines, both the weight and the volume of the baggage are utilized
to a maximum.
Further readings on linear programming can be found in [16].
18
18
16
14
weight = 8kg
knäckebröd, k 12
10
8 volume = 60l
6
4
2
0
0 2 4 6 8 10 12
messmör, m
Figure 3.2: Optimal solution to the problem of selling Knäckebröd

and Messmör. The optimal solution is K ≈ 7.6 and M ≈ 5.3.
3.2 Integer linear programming
An integer linear program (ILP) is formulated in the same manner as a linear

program with the additional restriction of the variables having to be integer
valued instead of continuous. If both integer and continuous variables are
used, the program is a mixed integer linear program (MILP). Though still
expressed in a linear way, solving a integer program is generally a more time
consuming task than solving its continuous counterpart. Moreover, changing
a problem from continuous to integer restricted means adding restrictions on
the variables, and additional restrictions can never increase the value of the
optimal solution but will decrease it or perhaps not change it at all.
To get a hint of why integer problems take a long time to solve, one could
compare to the complete enumeration of all possible combinations. If hav-
ing a set of n binary variables, xi ∈ {0, 1}, where i = 1, ..., n, the number
of possible ways to combine these variable values is 2n . With a computer
capable of computing and checking feasibility of one billion, 109 , of these
combinations per second, the time needed for different number of variables
is shown in Table 3.2. If the problem is that of connecting n wind turbines
by roads, similar to the formulation derived in Section 4.3.2, the number of
integer variables correspond to n + n2 , and the complete enumeration time,
using the same computer as above, is also seen in the table.
19
n time, 2n
20 0.001 seconds
50 313 hours
80 38 years
2
n time, 2n+n
5 1 second
6 1.2 hours
7 2.3 years
8 150 thousand years
Table 3.2: Computational time. Time required for complete enumeration

on a computer capable of 109 computations or feasibility checks per second.
Clearly, other means of methods than complete enumeration are required.

One commonly used algorithm, which will be used in this thesis, is the branch
and bound algorithm. This algorithm branches on the fractional variables
found as optimal solutions of the linear program (LP) relaxation of an integer
problem (say x?LP = 5.4 being the optimal solution of a LP-relaxed problem
where one of the integer variables is denoted by x), one at a time, splitting the
problem into two continuous subproblems where the variable is forced to be
no larger than the fractional value rounded downwards to the closest integer
(x ≤ 5) and no less than this value plus one (x ≥ 6), respectively. Each of
these problems is a LP-relaxation, meaning that the integer requirements are
relaxed. Fractional variables in the solutions of the two new problems are
branched on, creating two new subproblems per subproblem, and the process
is repeated. These iterations continue on a branch until an integer solution is
found, in which there are no fractional variables left to split in that branch, or
the bounds on the variables renders the problem infeasible. However, since
the solution to every subproblem will be an upper bound on the objective
value of that branch (remembering this value cannot increase by imposing
more restrictions), the algorithm will also stop searching a branch if this
upper bound is less than the objective of an already found integer solution.
This procedure is called pruning the solution tree. Throughout the branch
and bound process, better integer solutions will be found and the upper LP-
relaxation bound will be decreased. The relative difference between these
is a measure of the difference between the value of the best integer solution
and that of the optimal solution. It is thus a quality measure of the best
feasible solution found. It is here referred to as the MIP-gap, defined for a
maximization problem as
20
Best LP-relaxation objective value found so far

MIP-gap = − 1.
Best integer objective value found so far
The principle of the branch and bound method is illustrated in Figure 3.4
for the integer problem example of exporting Knäckebröd and Messmör, and
is described below.
Continuing the example problem 3.1, we may sell fractions of packages, but
when buying our goods in Sweden we will have to buy full packages. This
translates into requiring the variables representing Knäckebröd and Messmör
being integer, that is K M ∈ N, where N = 0, 1, 2 ... which is altered in the
mathematical formulation to
maximize 20K + 25M,

(3.2)
7.5K + 0.6M ≤ 60,
K, M ∈ N.
The graphical representation of the set of feasible solutions to this integer

problem is shown in Figure 3.3. For example, rounding the obtained contin-
uous optimum to the closest integer feasible point is not a way of optimizing
the integer problem.
Now, to solve this problem the branch and bound algorithm is used. The
procedure is hereby described and is illustrated in Figure 3.4. We start
off with the optimal solution to problem 3.1, (K, M ) = (7.6, 5.3). Now,
we choose a fractional variable, in this case K, and branch in to the two
continuous subproblems with the additional constraints K ≤ 7 and K ≥ 8,
respectively. When solving these continuous problems we notice an integer
solution, (8, 0) = 160, found in the "K ≥ 8"-branch. This is a solution to
the integer program 3.2, but we do not know whether it is optimal. The
MIP-gap tells that the value of the real optimum cannot be more than 75%
better than this integer solution. This branch can now be pruned and the
objective value of 160 is used as a lower bound on the optimal objective value.
If another branch yields an upper bound (a LP-relaxation value) lower than
this, that branch can be pruned. Continue by branching on the fractional
value of M in the leftmost node in level 1, yielding an even better integer
solution, (7, 5) = 265, and another fractional solution, respectively. From
21
18
16
14
weight = 8kg
12
knäckebröd, k
10
8 volume = 60l
6
4
2
0
0 2 4 6 8 10 12
messmör, m
Figure 3.3: The feasible set of solutions to the integer problem. The
dots all represent possible solutions.
this node we branch into K ≤ 6 and K ≥ 7, but due to the earlier constraint
of K ≤ 7 the only option for the rightmost branch is K = 7. Now, the
configuration of M ≥ 6 and K = 7 is infeasible and searching along that
branch is stopped. We carry on, find an even better integer solution, (6,6),
with an objective value 270, but still does not know whether it is optimal.
The next split however yields one infeasible subproblem and one for which
the upper bound is less than the objective of the best integer solution found,
and we have proven the solution (K, M ) = (6, 6) to be optimal.
The feasible set of solutions to this problem is once again, along with this
computed optimum, illustrated in Figure 3.5.
3.3 Graphs and trees
A graph, G = (V, E), is a set of nodes or vertices, V , linked in some way

by directed and/or undirectededges, E. An example of a graph is a set
of cities with air traffic routes linking them together either directly or via
another city in the set (but without requiring the existence of paths between
any pair of nodes). By representing the nodes and edges with variables (for
example, xi ∈ {0, 1}, i ∈ V , and yij ∈ {0, 1}, (i, j) ∈ E, respectively (see
22
(7.6, 5.3)
Level 0 MIP gap
283
K≤7 K≥8
(7, 5.6) (8, 0) 280.6

Level 1 280.6 160 −1 = 75.3%
160
M≤5 M≥6
(7, 5) (6.4, 6) €
278 278
Level 2 265 −1 = 4.9%
K≤6 K=7 265
(6, 6.25) 276.3

infeasable € −1 = 4.3%
Level 3 276.3
265
M=6 M≥7
€ 271
(6, 6) (4.8, 7) −1 = 0.4%
Level 4 270 271 270
K≤4 K≥5
€
(4, 7.5)
Level 5 infeasable
267.5
Figure 3.4: Branch and bound solution tree, for the problem of exporting
Knäckebröd, K, and Messmör, M.
23
18
16
14
weight = 8kg
12
knäckebröd, k
10
8 volume = 60l
6
4
2
0
0 2 4 6 8 10 12
messmör, m
Figure 3.5: Optimal solution to the problem of selling Knäckebröd

and Messmör. The optimal solution is K = 6 and M = 6.
illustration in Figure 3.6)), these roads connecting cities can be modeled by

the use of constraints in a linear program. It is thus, for example,
P possible
to minimize the total sum of the air route lengths (minimize (i,j)∈E dij yij ,
where dij is the length of the route from node i to j), while requiring each
visited
P city to be connectedPby both incoming and outgoing routes (subject
to i yij ≥ xj , j ∈ V and j yij ≥ xi , i ∈ V ).
2
1 3
4 5
y6,1
6 7
Figure 3.6: Graph example. Consisting of geographically distributed nodes

with directed edges connecting pairs of nodes.
There are numerous optimization problems formulated using graphs, as, for
example, the traveling salesperson problem, where one needs to find the
shortest path for a salesperson visiting every city in a set but never visiting
24
the same city twice, or as similar to the problems of this thesis, to find low
cost tree structures spanning the nodes. A minimum spanning tree, for ex-
ample, is to find the shortest, or cheapest, combination of edges that connect
each node to at least one of the other nodes, creating a structure of edges sim-
ilar to the branches of a tree. For many of the classic graph theory problems,
heuristic solution algorithms have been developed. However, when altering
such a problem or creating a somewhat different type of graph problem, as
in this thesis, these heuristic algorithms need not be applicable. Finding effi-
cient algorithms producing at least close to optimal solutions using heuristics
developed for a new problem could prove to be hard.
25
26
CHAPTER 4. FORMULATING THE OPTIMIZATION MODEL
Chapter 4
Formulating the optimization

model
27
4.1 What to formulate
As of today, the few wind farm layout optimization softwares available on

the market are primarily based on heuristic algorithms trying to find good
enough solutions in reasonable time [7, 17, 12, 23]. These implementations,
however, cannot provide a guarantee on the solution quality which a MILP
formulation can. There are thus incitements for formulating a MILP and
analyzing the outcome on realistic project areas.
The problems defined in Chapter 2 are addressed throughout this Chapter.

The first problem is that of maximizing the overall production, considering
the maximum number of turbines, minimum distances between turbines and
negative interference between turbines, so called wake effects. The second
problem deals with the construction costs of the park, for example, the foun-
dation cost at each possible turbine position and the costs of cables and roads
connecting the turbines.
4.2 Placing wind turbines and maximizing pro-

duction
4.2.1 Possible production at different locations
The production maximization problem is formulated on a graph G=(V, E)

where the set of nodes V defines possible wind turbine locations, and the
edges E denote the paths between each pair of nodes. The placement of
a wind turbine in a node in V is represented by the binary variable xi ,
which attains the value of 1 if a wind turbine is located at node i, and 0
otherwise. The objective is to maximize the sum of the production, pi MWh,
in every active node (i.e., a node i such that xi = 1) with an additional
constraint restricting the maximum number of installed wind turbines to n.
This constraint could also represent a restriction of the maximum installed
capacity. The integer linear program is thus to
28
X
maximize p i xi ,
i∈V
X
subject to xi ≤ n,
i∈V
xi ∈ {0, 1}, i ∈ V.
Multiple wind turbine models, or the option of choosing a lower production

level in a wind turbine, for example, to decrease sound levels, can be formu-
lated by expanding the nodes in V with new sets of xli , where l ∈ L denotes
the different types of wind turbines. Only one type in each location will
be chosen due to the constraints formulated in the next section. A maxi-
mum installed capacity constraint when using multiple wind turbines would
have
P to Pbe formulated as the slower method restricting the installed capacity,
l l
i∈V l∈types pi xi ≤ pmax , where pi is the wind turbine capacity for turbine
type l at location i, instead of the number of wind turbines.
4.2.2 Minimum distance between wind turbines
To prevent wind turbines from being placed too close to each other, con-
straints regarding the minimum separation distance between wind turbines
are formulated. This is done by stating that none of the other nodes within
a given range from an active node is allowed to be active. This minimum
separation distance is defined as a factor α times the rotor radius, R, when
modeling a circular minimum separation zone, but it can also be expressed
by for instance an ellipse with it’s major axis aligned with the primal wind
direction.
Setting the sum of all nodes within the minimum separation distance αR of
node i less than or equal to one will not do the trick since the constraint
has to hold also for a node where no turbine is placed. At this inactive
node it is perfectly alright for more than one node within the distance αR
to be active without them being less than αR meters apart, as illustrated in
Figure 4.2. In a continuous representation of possible locations this would
lead to a formulation where a maximum of five nodes strictly less than the
minimum separation distance from i would be allowed when node i is not
active, and one otherwise. Six active nodes around an active or non active
node is possible only when the distances from these to i is larger than or
equal to αR, shown in Figure 4.1. The mathematical representation of this
29
constraint will be the sum of five times the value of node xi plus the values
of the surrounding nodes xj , where j ∈ Qi , should be less than or equal to
five. The set Qi is defined as all the nodes j with distance from i less than
αR, except for the node i itself. |Eij | denotes the distance in meters between
the nodes i and j. The constraints read
X
5xi + xj ≤ 5, i ∈ V,
j∈Qi

where Qi = j ∈ V |Eij | < αR \ {i} , i ∈ V . The solution of this problem
is however slow and a different approach is therefore adopted. Here, the sum
of active nodes within a certain radius around a node is set to be less than
or equal to one. The radius used is of half the minimum separation distance
and now the program rely on the points in between each pair of nodes to
set the minimum separation distances, as illustrated in Figure 4.4. At most
one wind turbine may be placed within each of these circles. The problem
becomes one of packing circles as dense as possible, around active and non
active nodes, as seen in Figure 4.3. To allow for additional temporary nodes,
used only for these constraints but not as possible wind turbine locations, a
new set of nodes and edges, Ḡ = (V̄ , Ē) where G ⊆ Ḡ, is introduced. The
new constraint, which now will be used in the formulating of this Production
model, is given by
X
xj ≤ 1, i ∈ V̄ , Qi = j ∈ V̄ |Ēij | < αR/2 .
j∈Qi
The constraint says that the sum of the active nodes in a set Qi may not be
greater than one, where Qi consists of all nodes j on distance less than αR/2
from node i (including i itself). |Ēij | denotes the distance in meters between
the nodes i and j.
In a discrete representation of possible locations situations can occur where

there does not exist any useful points in between two nodes, allowing them
to be placed within the minimum distance of one another. This is easily
corrected by defining the added temporary nodes at half the distance between
each pair of nodes. Obviously this is required only for pairs of nodes separated
by approximately the minimum distance, or more specifically by more than
half but no more than the full minimum separation distance. Creating these
30
Figure 4.1: Six wind turbines in an area less than the minimum
separation distance of a node is not possible. The tightest placement
of the six are instead on the minimum separation distance.
Figure 4.2: Strictly less than the minimum separation distance of

a node, a maximum of five wind turbines can be placed, depending on the
status of the center node.
31
Figure 4.3: By packing circles of half the minimum separation dis-

tance radius, a maximum of five wind turbines strictly in the area less
than the minimum distance of a node is achieved. The locations of the wind
turbines are the same as in Figure 4.2.
Figure 4.4: A maximum of one active node in the circle area less
than half the minimum separation distance of a node, forces two
active nodes to be separated by at least the minimum separation distance.
32
sets is particularly easily done in a rectangular or diamond grid, where one

just substitutes the grid for that of twice the grid resolution. Minimum
separation distance constraints are then created with the nodes in V̄ , after
which the variables of the temporary nodes are deleted but the constraints
retained.
4.2.3 Packing density in a discrete grid
To allow for close to hexagonal packaging, a grid size of the minimum sepa-
ration distance divided by eight or multiples of eight is recommended. On a
bounded area, for which a hexagonal or a regularly spaced packing is optimal,
optimization on a discrete grid may yield a solution close to the continuous
optimum. Examples of this are shown in Figure 4.5, where the proven con-
tinuous optima are obtained from [11] and [10].
4.2.4 Production loss between wind turbines
To account for wake effects, that is the loss of production when placing two
wind turbines close to each other (see Section 2.1.4), a new variable wij ∈
{0, 1} is introduced. This variable, if equal to one, represents a production
reduction at node j due to a wind turbine at node i. This constraint equals
the one of the generalized vertex packaging problem [20] and is also used
by Donovan in [4, 5] and Donovan et al. in [6]. The magnitude of the
corresponding production loss is denoted dij MWh and must be calculated for
every pair of nodes since it depends on the wind frequency and distribution
for the specific location and direction of wake. Since the proximity range
constraints of Section 4.2.2 prevent the use of edges shorter than αR, these
can be excluded from the set E, defining the subset Ê ⊆ E. The modified
objective function is to maximize
X X
p i xi − dij wij ,
i∈V (i,j)∈Ê
and the added constraints are
33
1800 1600
1600 1400
1400 1200
1200
1000
1000
800
800
600
600
400
400
200
200
0
0
−200
−200
0 500 1000 1500 0 500 1000 1500
[meter] [meter]
(a) 19 circles (b) 20 circles
1800
1600
1400
1200
1000
800
600
400
200
−200
0 500 1000 1500
[meter]
(c) 25 circles
Figure 4.5: Circle packing. Placement of 19, 20, and 25 wind turbines,
respectively, in different geometric areas. The inner black dashed line rep-
resents the area required for packing the number of circles in a continuous
manner, whereas the outer dashed line is the area required using the for-
mulated linear program. The two dashed lines differ little or nothing in the
three examples. The solid line is the project area used by the linear program.
34
xi + xj − wij ≤ 1, (i, j) ∈ Ê,

wij ∈ [0, 1], (i, j) ∈ Ê.
The variables wij ∈ Ê may take only binary values, but due to the formula-
tion of the constraint relating two nodes and their edges this requirement will
be fulfilled in an optimal solution alsoP without integrality requirements on
wij . The sum of the losses from wakes, (i,j)∈Ê dij wij , is subtracted from the
objective function, and the wake constraint forces wij to 1 if both xi = xj = 1
(i.e., the nodes i and j are active), otherwise it will take the value of 0.
The formulation of wake effects between two possible locations using linear
relations is not restricted to any particular wake model. However, due to
linearity the combination of multiple wake effects on a node equal the super-
position of all wake effects from active locations on this node. If non-linear
superposition is preferred, approximations by linearizations are addressed in
Section 4.2.7.
4.2.5 Maximum sound levels
To deal correctly with maximum allowed sound levels one cannot simply de-
fine a minimum allowed distance to wind turbines from a noise sensitive area,
denoted by an index s ∈ S̄, corresponding to a certain sound level. This is
because the sound pressure from two wind turbines will be superposed and
the minimum allowed distance thus increased. The constraint regarding max-
imum sound levels is formulated as the superposition of the individual sound
pressures ais in s from active nodes i, which must not exceed a maximum
allowed value, amax . When using multiple types of wind turbines, their differ-
ent sound characteristics must also be implemented in this constraint. Sound
pressure (SP) and sound level (SL) relate as SL = 10 log10 (SP). The resulting
constraint is given by
P
i∈V ais xi ≤ amax , s ∈ S̄.
35
4.2.6 The complete Production model
Combining the formulations in Sections 4.2.1, 4.2.2, 4.2.4 and 4.2.5 yields
the complete Production model, which is here described and presented.
The Production model optimizes the total production when placing wind
turbines in a project area. It can place at maximum a given number, n, or as
many wind turbines as possible. The model respects the minimum separation
distance between wind turbines, αR, and that the sum of the individual sound
pressures, ais , originating from the wind turbines, is held below the maximum
level, amax , for any noise sensitive area, as . The production, pi , in each
possible location, and the production loss, dij , between wind turbines i and j
due to wake effects, are calculated from any production and wake effect model
(see Sections 2.1.2 and 2.1.4), respectively. The graph G = (V, E) consist of
all the nodes V and the edges E connecting each pair of nodes. Ĝ = (V̂ , Ê) is
the subgraph of G with edges shorter than the minimum separation distance,
αR, removed, and Ḡ = (V̄ , Ē) is the graph with twice the grid resolution of
G, used for implementing the minimum separation distances. The complete
Production model is to
X X
maximize p i xi − dij wij ,
i∈V
X(i,j)∈Ê
subject to xi ≤ n,
i∈V
X
xj ≤ 1, Qi = {j ∈ V̄ |Ēij | < αR/2}, i ∈ V̄ ,
j∈Qi
xj − wij ≤ 1,
xi + X (i, j) ∈ Ê,
ais xi ≤ amax , s ∈ S̄,
i∈V
xi ∈ {0, 1}, i ∈ V,
wij ∈ [0, 1], (i, j) ∈ Ê.
The model can also, through defining a revenue on production, be shifted

to that of optimizing profit. Here, locally varying costs of foundations can
also be incorporated. Additional formulations reducing the number of wake
variables, reducing the MIP-gap and addressing linearization of non-linear
combination of wake effects are described in Section 4.2.7, however not im-
plemented.
36
4.2.7 Additional formulations
These suggested formulations or ideas have not, due to time restrictions, been
implemented in this thesis. They seem however well worth to investigate in
future work.
Dividing an area into sectors to reduce the number of wake vairables
A high resolution grid can be of importance in the tight packing of circles.

The accuracy and level of detail of the wake effects need not however depend
as greatly on this resolution. It would then mean a great improvement of
problem size if the number of wake variables, wij , did not increase by the
square of the number of nodes. The idea is to divide the entire area into
square sectors
√ with side lengths less than the minimum separation distance,
αR, by 2 to contain a maximum of one active node. The number of possible
locations within a sector could be large, but the wake effects are only to be
calculated between active sectors, see Figure 4.6. Thus, the numbers of wake
variables and corresponding constraints grow by a factor of |V | instead of
|V |2 . Now, moving two wind turbines relative to one another will affect
the production in the locations picked, as well as their internal distance
and bearing which wakes are calculated from. Examples of distributions
in relative movement and bearing between two wind turbines when shifting
their location to a sector centre is shown in Figures 4.7 and 4.8. The level
of approximation seem reasonable and can be adjusted by shifting the sector
side length. The modified wake constraint, if denoting the graph of sectors
and interconnecting edges by Gs = (S, E s ), is given by
X X
xk + xk − 1 ≤ wij , (i, j) ∈ Eijs ,
k∈Si k∈Sj
where Si ⊆ V , i = 1, ..., |S|, is the subset of the nodes within square sector
i. The set E s are composed by the edges connecting each pair of sectors.
37
Figure 4.6: Wakes are calculated between active sector centers,

whereas the resolution within a sector could be much higher.
14
12
Cumula&ve distribu&on [%]
10
8
6
4
2
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Rela&ve movement [%]
Figure 4.7: Relative movement, showing the cumulative distribution of

the change in distance between two randomly placed wind turbines when
shifting their locations to the nearest respective sector centers. One wind
turbine is placed in the square sector with centre in the point (x, y) = (0, 0),
the other within x, y ∈ [−2000, 2000], respecting the minimum separation
distance of 400m. The sector side length is 100m.
38
0,14
0,12
Cumula&ve distribu&on [%]
0,1
0,08
0,06
0,04
0,02
0
0 1 2 3 4 5 6 7 8 9 10
Angular shi8 [deg]
Figure 4.8: Angular shift, showing the cumulative distribution of the shift
in bearing between two randomly placed wind turbines when shifting their
locations to the sector centers. One wind turbine is placed in the square sector
with centre in the point (x, y) = (0, 0), the other within x, y ∈ [−2000, 2000],
respecting the minimum separation distance of 400m. The sector side length
is 100m.
Reducing the MIP-gap by constraining the LP relaxation
In an optimal solution to the LP-relaxation of the production problem typ-

ically xi ≤ 0.5, i ∈ V , which implies that there are no active wakes be-
tween any nodes. Assuming a minimum allowed number of wind turbines,
K, preferably the number of wind turbines used as the maximum number
(K = n), the following constraints can enforce the proper total amount of
wake effects to be activated and thus reduce the MIP-gap. The effect on
the solution time due to the increased number of constraints has not been
investigated. Using K as the minimum number of wind turbines, constraints
forcing the total amount of incoming and outgoing wakes at an active node
are given by
X
wij ≥ (K − 1)xi , i ∈ V̄ ,
j∈E
X
wij ≥ (K − 1)xj , j ∈ V̄ .
i∈E
These constraints do, however, only force one of the nodes to which the wake
39
is connecting being active, and the use of small wakes to and from non-active
nodes at a far distance is possible. Additional constraints guarantee that the
nodes that an active wake is connecting to are also active, according to
wij ≤ xi , i, j ∈ V̄ ,
wij ≤ xj , i, j ∈ V̄ .
These four inequalities imply that the number of active wakesPrelatesP cor-
rectly to the minimum allowed number of active nodes, i.e., i j wij =
K(K − 1). This does, however, not force the LP-relaxation to take on in-
teger values. It is, for example, possible to split two quite separated wind
turbines into two times two 0.5-valued ones next to each other. This allows
for two times 0.5 valued wakes leaving each half wind turbine, resulting in
a total wake value of two instead of one between the original wind turbine
locations. The LP relaxation will by this splitting behavior yield large sums
of wake variable values between distant nodes, and less between nodes close
to each other.
Approximating other ways of combining wake effects
A linear superposition of wake effects will overestimate the combined produc-

tion losses compared to, for example, the ’sum of squared velocity deficits’-
method [14, 22, 9]. To compensate for this difference in a linear program, a
positive contribution to the energy production would have to be added when
a wind turbine is affected by two or more wakes simultaneously. To com-
pensate for every possible combination of wakes affecting a wind turbine, too
many variables would have to be added for the model to remain of reasonable
size. However, if accepting the added positive production to be an average
of the effect of, say, two arbitrary placed wind turbines placed within 1000m
north of the affected wind turbine instead of the positive contribution from
their exact locations, a somewhat simpler model can be formulated. Let the
n
binary variable Ki,θ represent n number of wakes in the angular segment θ
of node i affecting a wind turbine in node i. The added production com-
pensation of this variable is pn+1 n
i,θ ≥ pi,θ , meaning the more wakes combined
the larger the production contribution. With N = 2, 3 ... being the possible
number of interfering wakes, and Θ the set of angular sectors around each
node, typically the full circle divided by the angular spread of the wakes, and
S(i, θ) being the set of nodes in the θ segment of node i, the problem is to
40
XXX
maximize pni,θ Ki,θ
n
,
i∈V
Xθ∈Θ n∈N X
n (4.1)
subject to Ki,θ ≤ wki , i ∈ V, θ ∈ Θ,
n∈N k∈S(i,θ)
n
Ki,θ ∈ {0, 1}.
The formulation requires a number of added integer variables; this number is

equal to |V | · |Θ| · |N |. This is very many additional integer variables, and at
the price of further approximations yet another approach is hereby proposed.
Instead of correcting for multiple wakes in a number of sectors around each
node, a positive production addition that is proportional to the total number
of wind turbines aligned in a certain direction of the wind is suggested. This
can be seen as sectors looking like to long stripes covering the project site.
Sectors in a number of directions will be added, which can be divided into
covering not the full stripe but a part of it. The formulation of the linear
n
program will then be similar to that of 4.1, but the choice of Kl,θ for stripe l
in angle θ will now depend on the number of active nodes within the sector
instead of the number of wakes affecting a certain node. This approximate
formulation is to
XXX
maximize pnl,θ Kl,θ
n
,
Xθ∈Θ n∈N
l∈L X
n (4.2)
subject to Kl,θ ≤ xi , l ∈ L, θ ∈ Θ,
n∈N i∈S(l,θ)
n
Kl,θ ∈ {0, 1}.
The number of additional variables in this model is |L|·|Θ|·|N |, where L is the

set of stripes, Θ the set of directions of the stripes (which in this formulation
only need to cover [0,π)) and N = 2, 3, ... is the possible number of wind
turbines within the stripe. This will generally result in a lot fewer variables
compared to that of the model (4.1) but it is a more rough approximation and
does not account for the wakes expanding. However, the two models (4.1)
and (4.2) provide an approximate basis for correcting the linear superposition
of wakes to more closely relate to a non-linear combination method.
41
4.3 Infrastructure formulation
Another optimization area of interest concerns the costs associated with the
constructions of foundations as well as roads and cables necessary for the wind
farm. In these formulations the locations of the wind turbines are considered
fixed. However, the simultaneous solution of the Production model and the
Infrastructure model, not requiring the locations to be fixed, is possible but
seems to be a rather complex problem.
4.3.1 Foundation costs
The foundation costs of Section 2.2.1, ci , are modeled simply as a negative

revenue per active node i, added to the objective function. The objective
function is then to
X
maximize − ci x i ,
i∈V
subject to xi ∈ {0, 1}, i ∈ V,
which, if the locations of the wind turbines are fixed by the prior solution of
the Production model, can only take on one value.
4.3.2 Formulating a Steiner tree
When connecting the wind turbine locations with roads and power lines, a
spanning tree structure, as defined in Chapter 3.3, is to be constructed. An
example of this is connecting the nodes of Figure 4.9 as shown in Figure 4.10,
a minimum spanning tree solution obtained by the model described below.
This example will be addressed again when discussing cables in Section 4.3.4.
Each edge is associated with a certain cost, dij defined by a number of fac-
tors such as length, topography and such. A linear programming formulation
which will minimize the sum of the edge costs is sought. However, the formu-
lation should also allow for, but not require, the tree to use paths involving
nodes that are not active as locations of wind turbines. Such a formulation
42
4500
4000
3500
3000
2500
2000
1500
1000
500
0
0 1000 2000 3000 4000 5000
Figure 4.9: Randomly scattered nodes, to be connected by the cheapest

possible tree structure. The bottom left square node is also to be connected,
but will later be used as the point of connection to the external grid.
19 10
4500 10 22
1
2
16 11 22 23
11 9 3
4000 9 8 23 16
8 13 7 2 23
1 13 9 3
3500 4 9 6 4 21
14 21 2 1
24 13
2
3000 20 24 2
1 6 17
15
2500 18 17 17 7
1 17
7 12
2000 18
1500
12 14
19
26 25 14 26
1000 23 22
25 3
24 15 14
500 2
3 1 5 15
25 1
0
0 1000 2000 3000 4000 5000
Figure 4.10: A minimum spanning tree considering only the layout of the
edges and not minding that they are directed, connecting the nodes of Figure
4.9. There are no limits on the flow through the edges. The upper two figures
of each triple of numbers state which two nodes the edge is connecting and
the bottom figure state the flow through the edge.
43
could allow for a shorter tree (if defining an euclidean tree) or ease the pro-
cess of formulating paths around obstacles (Figure 4.11), and results in a so
called Steiner tree problem [3, 2]. Since the problem is that of electricity
flow, at some central point of the tree the flow should be passed on to an
external grid.
Figure 4.11: Shorter paths are possible, when using other nodes than
just the active ones (Steiner tree formulation). The left path is 5% longer
than the right one.
A positive revenue, ri , per node xi ∈ V , is defined such that the solution to

the minimization of tree costs does not lead to only the trivial solution of
all xi being zero. When using this tree formulation in conjunction with the
Production model, this revenue is replaced with the one from that model. The
nodes with positive revenue are the ones previously selected as wind turbine
locations. The graph of possible nodes and edges for the tree formulation
is defined as Gt = (Vt , Et ), where the set V from the Production model is a
subset of Vt and Et are the edges connecting each pair of nodes in Vt . The
objective function of the tree formulation of this graph, using the binary
variable zij ∈ Et to define the use of an edge from i to j, then is to
X X
maximize (ri − ci )xi − dij zij , (4.3)
i∈V (i,j)∈Et
where xi ∈ {0, 1}, i ∈ V and zij ∈ {0, 1}, (i, j) ∈ Et .
To deal with the actual creation of the tree structure, and to prevent the
occurrence of subtours, a new continuous variable yij ∈ [0, f ], representing
the flow of electricity in the tree, is introduced. Here, f ∈ [0, n], where
n is the maximum allowed number of nodes connected at, or flow in, edge
(i, j). Each active node creates one unit of flow and flow entering a node
must also leave the node. The accumulated flow at the connection node (the
external power grid), denoted CN ∈ Vt , must equal at least the number
44
of active nodes minus one. We denote the flow on edge (i, j) by yij . This
connection node constraint is not really necessary, since the flow must be
exported somewhere and the connection node is the only option, but it was
discovered having a slight effect on reducing the time of finding a feasible
solution. The constraints read
X X
yij + xj − yjk = 0, j ∈ Vt \{CN },
ij∈Et jk∈Et X
(4.4)
P
i∈Vt yi,CN ≥ xi − 1,
i∈V
yij ∈ [0, fij ], (i, j) ∈ Et .
Now, yij is the number of wind turbines connected on the branch from node
i, including i, and it’s flow can be limited either individually for each edge
(i, j) or for all (i, j) ∈ Et . The binary variable defining the activation of
an edge is however zij and to induce the activation and cost of an edge we
connect these two by replacing the last line of (4.4) by the constraints
0 ≤ yij ≤ fij zij , (i, j) ∈ Et . (4.5)
Using the formulation illustrated in (4.4) with the modification of (4.5), iso-
lated subtours are not possible due to the accumulative property of the vari-
ables yij , illustrated in example Figure 4.12. There exists feasible solutions
with subtours connected to paths leading to the connection node, but it is
then cheaper to break up the subtour, as illustrated in Figure 4.13.
3+1
0+1
1+1
2+
1
Figure 4.12: Isolated subtours, are not possible since flow is added in each
active node and the flow along an edge can take on only one value.
The basic tree formulation can then be written as the linear program to
45
3 3
1
3 2
Figure 4.13: Non-isolated subtours, are possible, but breaking the subtour
will yield a shorter path.
X X
maximize ri xi − dij zij ,
i∈V (i,j)∈Et
X X
subject to yij + xj − yjk = 0, j ∈ Vt \{CN },
i:(i,j)∈Et k:(j,k)∈Et
X X
yi,CN ≥ xi − 1,
i∈Vt i∈V
0 ≤ yij ≤ fij zij , (i, j) ∈ Et ,
xi ∈ {0, 1}, i ∈ V,
zij ∈ {0, 1}, (i, j) ∈ Et .
On top of this formulation, optional constraints are added. The following

constraints do not alter the optimum solution, but have proven to reduce
computation time (in the sense of improved solutions after any tested amount
of time). The constraints
yij ≥ zij , (i, j) ∈ Et ,
relates the flow through and activation of an edge in the opposite way that
(4.5) does. The need for at least one active edge variable, zij , per active node
is expressed by
X
xi − zij ≤ 0, i ∈ V.
j:(i,j)∈Et
46
A flow leaving a node requires either flow entering the node or the node itself
being active, which is formulated by the constraint
X X
zij + xj − zjk ≥ 0, j ∈ Vt \{CN }.
Moreover, forcing flow in only one direction per edge is beneficial and can be
expressed by
zij + zji ≤ 1, (i, j) ∈ Et .
However, flow in multiple directions along an edge should be permitted but

will be prohibited elsewhere and this latter formulation is therefore not an
restriction. More on this in Section 4.3.7.
4.3.3 Minimizing the costs of roads
It is possible to limit the number of edges per node in a grid to only the
shortest edge in, say, sixteen directions, resulting in a bit more costly span-
ning tree but reducing the number of variables zij and yij to 16|Vt | instead
of |Vt |2 each. If roads are to be connected along an existing road instead
of at a single defined point, nodes with corresponding edges could be added
along this road and the cost of these edges set to zero, as in the example
illustration of Figure 4.15. Since roads and cables typically follow the same
paths – due to the need of a road to place a cable – the cost of a road is
merely an additional cost added to the cable cost, defined in Section 4.3.4,
and no additional road formulations to the Steiner tree are thus given here.
Two cables along the same path only requires one road to be built, which
is solved in the same manner as for a single cable by defining double cables
as a cable type in the mathematical formulation. This implemented model
will be able to handle double cables when flows are parallel. However, in
reality two cables with opposite flows will also be beneficial in terms of road
costs, as per the rightmost illustration in Figure 4.14. Another formulation
addressing this issue, where separate road variables are introduced, is given
in Section 4.3.7, however not implemented.
47
1 1
1
1 2
2
2
Figure 4.14: Three different ways of connecting the wind turbines. 1)

The shortest spanning tree using only edges between wind turbine locations.
2) The cost of combining cables is low relative to the standard cable cost.
3) Cost of cable is low relative to the cost of combining cables and multiple
cables on the same path yields only one road cost.
Figure 4.15: Using existing roads may reduce the cost of the path. Costs
are given in the diamond symbols. The cable path following the existing road
is longer but cheaper than the straight path.
48
However, if the tree of roads is to be connected at a different point than the

connection node, where the electrical flow is exported, additional constraints
are introduced. These constraints requires that at least one node in the
spanning tree of edges – not necessarily a location of a wind turbine – is
connected to any of K possible road connection points. The formulation
starts by defining a cost Cki for connecting point k ∈ K to node i. The
connection variable, Rki , is forced to zero if no flow is sent from the node i
through any of the edges (i, j), where j ∈ Vt , meaning that node i is not a
part of the tree. With this variable connecting the external road points only
to nodes being part of the Steiner tree, all that is left is to require at least
one of the variables Rki corresponding to these connection roads to take the
value of one, and thus being active. This linear program formulation is to
XX
maximize − Cki Rki ,
X i∈Vt k X
subject to zij − Rki ≥ 0, i ∈ Vt ,
j∈Vt X k∈K
X
Rki ≥ 1,
i∈Vt k∈K
Rki ∈ {0, 1}, i ∈ Vt , k ∈ K.
4.3.4 Cables
The capacity of a cable yij is upper bounded by the maximum number of

connected wind turbines possible for that type of cable, fij . The node con-
figuration of Figure 4.9, with only wind turbine locations as possible edge
connection nodes, is solved with an upper limit of the number of wind tur-
bines per edge, except for the unlimited export cable to node 1. The resulting
layout is shown for a limit of three wind turbines per edge in Figure 4.16 and
seven in Figure 4.17. Multiple cable types, l = 1, 2, ..., with different maxi-
mum capacity and costs, are introduced as the sets of variables yijl ∈ [0, fijl ]
and zijl ∈ {0, 1}, corresponding to the same edges (i, j) as for the original
cable type. The cable types may include, for example, larger dimension ca-
bles or double low-capacity cables. Including these new variables in the flow
forwarding and connection node constraints (4.4) will allow the model to
choose between the cable types. The higher price of the high capacity cable
will result in a solution that uses the cheapest cable that meets the capacity
requirements. The previous example using both cable types mentioned, with
49
different costs, is shown in Figure 4.18. Whether the solution time will ben-
efit from an additional constraint restricting the use of only one zijl per edge
(i, j), which in practice will always be the case, has not been investigated.
19 10
4500 10 22
1
2
11 9
4000 1 23 16
8 13 1
1
3500 9 6 4 21
2 21 2 1
16 7 2
3000 20 24 13 17 2
1 2
22 26
24 18 3
2500 2
6 26 2 26
3 3
2000 17 26
26 7 26
18
33 3
1500 12 26
1
25 26 14 26
1000 2 3
3 25
126 1
15 14
500 25 2 5 15
1
0
0 1000 2000 3000 4000 5000
Figure 4.16: An upper limit of three wind turbines per edge. With a
cable cost of 0.6 MSEK/km, and export cable cost of 3 MSEK/km, the total
cost of connecting the nodes of Figure 4.9 is 22.7 MSEK. This configuration
was found after about 25 s with a MIP-gap of 4% and was proven optimal
after 243 s. The cable connecting to the external power grid in point (0,0) is
the high capacity export cable originating from the transformer station.
Now, the transition between two cable types will occur either when passing a
wind turbine, thus shifting the capacity demand, or when combining the flow
from two different locations into one cable. The former case as well as a split
at a wind turbine location is considered to come without extra costs, since the
necessary equipment is already present. However, when the cables are to be
combined at a location other than at a wind turbine, the cost of the required
substation has to be accounted for. This is formulated by introducing a
binary substation variable, sj , where j ∈ Vt , with the associated cost gj ,
which should be forced to a value of one if flows are combined outside of
wind turbine locations. If the combination occurs at the location of a wind
turbine, sj = 0. This is achieved by requiring either sj or xj to take on a
value of one if more than one flow enters the node j. The flow combination
program is thus to
50
19 10
4500 10 22
1
2
11 9 16 11
4000 2 23 16 22 4
3 3
8 13 1
1
3500 9 6
4 4 21
13 24 21 2 4
5
3000 20 24 2 2 7
1 6 17
5 6
2500 17 18
6
2000
24 3
4 7 3
18 3
1500 7 7
12 14
1
26 25 14 26
1000 5 4
25 3
6
15 14
500 2
3 1 5 15
25 1
0
0 1000 2000 3000 4000 5000
Figure 4.17: An upper limit of seven wind turbines per edge. With a
cable cost of 1 MSEK/km, and export cable cost of 3 MSEK/km, the total
cost of connecting the nodes of Figure 4.9 is 24.3 MSEK. This configuration
was found and proven optimal after 11 s. The cable connecting to the external
power grid in point (0,0) is the high capacity export cable originating from
the transformer station.
X
maximize − gj sj ,
X j∈Vt
subject to zij − 1 − M (xj + sj ) ≤ 0, j ∈ Vt , (4.6)
i∈Vt
sj ∈ {0, 1}, j ∈ Vt ,
where M equals the maximum number of incoming flows to a substation.

It would be possible to relax the integrality requirement on sj , and rewrite
−M (xj + sj ) as −(M xj + sj ), with an upper limit of sj higher than one,
but the cost of the substation would then be proportional to the number of
incoming flows, which is typically not the case.
Flows can be combined, but are not permitted to be split, which calls for
additional constraints. These constraints, limiting the number of flows out
of a node i, are defined as
X
zij ≤ 1, i ∈ Vt .
j∈Vt
51
19 10
4500 10 22
1
2
22 23
11 9 3
4000 1 23 16
8 13 4
1 9 13
3500 3 6 9
1 4 21
13 24 21 2 1
16 17 2
3000 20 24 5 5 2 7
1 3
2500 17 18
6
7 12
2000 4
24 3
7
18 3
1500 7 12 26
5
26 25
1000
25 3 14 6
3
7 3
15 14
500 2
3 1 5 15
25 1
0
0 1000 2000 3000 4000 5000
Figure 4.18: Multiple cable variables with upper limits of three and
seven connected wind turbines, respectively. Using the prices from
Figures 4.16 and 4.17 the total cost of connecting the nodes of Figure 4.9
is 20.4 MSEK. This configuration was found after about 50 s with a MIP-
gap of 6% and was proven optimal after 47 min. The cable connecting to the
external power grid in point (0,0) is the high capacity export cable originating
from the transformer station.
52
4.3.5 Transformer station placement
The wind turbines of a wind farm generally need to be connected to the

export cable through a transformer station. In Figures 4.16 – 4.18 the loca-
tion of the transformer station is where the red export cable connects to the
tree. Its optimal location will depend on the length and cost of the internal
cables, the export cable and the foundation costs at the given location. If
the latter does not vary locally, the cost of the transformer station could
potentially be corrected for later, since the location of the export cable con-
necting to the root of the tree then only will depend on the relative prices
of the cables. However, the use of transformer placement variables and costs
has the additional function of restricting the number of transformer stations.
Without these costs it could, for example, be beneficial to place two export
cables at different locations and thus constructing two trees with the require-
ment of two transformer stations. This problem could of course be solved
in other
P ways, for example, by restricting the number of export cables to
one, i∈Vt zi,CN ≤ 1, but for this model the transformer placement, includ-
ing locally varying costs, is included. The objective function is extended by
the sum of the costs ctr i of the continuous transformer placement variables
ti ∈ Vt \{CN }, which will take on the values only 0 or 1 due to the fact
of zi,CN being binary. The linear programming formulation of the added
transformer costs is to
X
maximize − ctr
i ti ,
i ∈Vt \{CN }
subject to zi,CN ≤ ti , i ∈ Vt ,
ti ∈ [0, 1].
4.3.6 The complete Infrastructure model
Combining the above formulations yields the complete Infrastructure model,

which is hereby described and presented.
The Infrastructure model is that of connecting the wind turbines by roads and
cables, and of locally varying foundation costs. If the locations of the wind
turbines are fixed, the location variables xi could be set to their defined values
and no revenue, ri , would be necessary. For the purpose of combining the
53
two models, however, this revenue is included in the objective function. The
graph Gt = (Vt , Et ) contains the nodes and edges used to form the Steiner
tree. The set, V , of possible or fixed locations of wind turbines is a subset
of Vt . The nodes of Vt can be many more than the number of wind turbines,
making cheaper connections possible, and includes the point of connection
to the external grid, xCN . Power flow along an edge is represented by yij
and one unit of flow is added in each active node along a cable. The sum of
the flow entering the connection node must not be less than the number of
active nodes, or this number minus one if wind turbine placement is allowed
in the connection node. Due to this accumulated flow and the need to export
the flow, isolated subtours are prohibited. Several types of cables, yijl , where
l = 1, 2, ..., with different maximum flows, fijl , can be used on an edge. The
cost, dlij , of cable and road type l on edge (i, j) is activated by the binary
variable zijl . Moreover, constraints requiring at least one road Rki with cost
Cki be connected to the road tree are included. There are also constraints
introducing a substation cost gj through the combination variable sj where
flow is combined outside a wind turbine location. Flow splits are prohibited
by another constraint, and the location specific cost of a transformer station,
ctr
i , is introduced at each transformer station location by the variable ti . The
complete Infrastructure model is to
54
X X XX X X
maximize r i xi − dij zij − Cki Rki − gj sj − ctr
i ti ,
i∈V (i,j)∈Et i∈Vt k j∈Vt i ∈Vt \{CN }
X X
subject to yij + xj − yjk = 0, j ∈ Vt \{CN },
X X
xi − 1 ≤ yi,CN ,
i∈V i∈Vt
0 ≤ yij ≤ fij zij , (i, j) ∈ Et ,
X yij ≥ zij , (i, j) ∈ Et ,
xi − zij ≤ 0, i ∈ V,
j:(i,j)∈Et
X X
zij + xj − zjk ≥ 0, j ∈ Vt \{CN },
i:(i,j)∈Et j:(j,k)∈Et
X X
zij − Rki ≥ 0, i ∈ Vt ,
j∈Vt X k∈K
X
Rki ≥ 1,
X i∈Vt k∈K
zij − 1 − M (xj + sj ) ≤ 0, j ∈ Vt ,
i∈Vt
zi,CN ≤ ti , i ∈ Vt ,
xi ∈ {0, 1}, i ∈ V,
zij ∈ {0, 1}, (i, j) ∈ Et ,
Rki ∈ {0, 1}, i ∈ Vt , k ∈ K
sj ∈ {0, 1}, j ∈ Vt
ti ∈ [0, 1], i ∈ Vt .
4.3.7 Additional formulations
Reducing the number of nodes when wind turbine positions are

held fixed
If turbine positions are held fixed the number of nodes in the grid defining
the tree structure can be reduced. It would, for example, generally not be of
any use to keep nodes on the outside of the "outermost" wind turbines. Also,
better geographic locations for combining flows, for example, at the centre of
a group of wind turbine locations, can be found for various combinations of
the wind turbines. Thus, the set of nodes can through a bit of smart prepro-
cessing be redefined to a smaller but better placed set of nodes. The grid can
55
(a) tree length: 1.03 (b) tree length: 1
(c) tree length: 1.08 (d) tree length: 1.15
Figure 4.19: Reduced number of nodes. A principal example where 48

nodes are reduced to 11 consisting of nodes half way between active nodes
and of nodes at a few cost weight centers. (c) and (d) only serve to illustrate
possible new paths.
hence be transformed into a set of locations which suits the configuration of

wind turbines even better. Figure 4.19 shows examples of these principles.
If the Infrastructure model is combined with the Production model, and thus,
the locations of the wind turbines are not fixed, this type of rearrangement
is not possible. However, a similar approach to that of defining wake effects
between sectors (Section 4.2.7) may be adopted, creating the tree structure
connecting geographical sectors instead of specific points (which are included
in these sectors).
Separation of tree variables into cables and construction variables
Two or more cables on the same path is in the Infrastructure model formu-
lated by an edge variable with the cost of multiple cables but the cost of
only one road. However, it cannot handle, for example, opposite directions
of flow, which would also lead to only one road having to be built. A sim-
ple solution to this problem, increasing the number of binary variables and
56
constraints by the number of undirected edges, is to split the active path

variable zij into cable variables zijc and road construction variables zijr . This
does not call for creating two sets of trees – the cable variables are used as
before in, for example, the flow forwarding constraints (4.4), but the cost of
road construction is moved to the variable zijr which is activated whenever
a cable path is used. Since the road construction cost does not depend on
the direction of the power flow, two nodes need to be connected by only the
undirected zijr , yielding the constraints
c
zijc + zji ≤ 2zijr , {j ∈ Vt | j > i}, i ∈ Vt ,
r
zij ∈ {0, 1}.
This solves the problem of the costs not decreasing when using a path as in
the rightmost illustration of Figure 4.14. However, due to the restriction of
not allowing splits of flow, this type of path, where flow enters and leaves
and then enters and leaves the same location again, is still not possible. A
solution to this is to also implement the idea in the next section.
Multiple nodes per location to allow for additional paths
Even though it is sometimes economically beneficial to use a path as in the

rightmost illustration of Figure 4.14, the restriction of not being able to split
flow in a node also makes it prohibited to split flow in the same location.
One solution to this problem would be to add an additional set of nodes in
the same locations as the original set. Multiple nodes in the same location
means that one flow could enter and leave one node and a second flow could
enter and leave another node but in the same geographic location. Using the
additional formulation above, with i1 representing node 1 at location i and l
being the number of nodes per location, the constraints read
X X
zicm ,jn + zjcn ,im ≤ M zijr , i, j ∈ Vt : j > i,
m={1,l} n={1,l}
where M ≥ l. This shift of nodes and edges requires changing the sets used
in many of the other constraints too.
The problem size increases rather fast by the introduction of these multiple
57
nodes, mainly because the number of edge variables yijl and zijl grows by the
square of this multiple. The number of variables needed in different locations
of the project area could however vary, as, for example, shown in Figure 4.20,
which would result in the problem size growing at a slower rate.
Figure 4.20: Multiple nodes in each location. The number of possible

nodes can be determined locally depending on the characteristics of the setup.
Now, this formulation is preferably used when the locations of the wind
turbines are fixed, whence multiple nodes are not needed at these locations,
and the rearrangement of nodes to geographically superior locations, as in
Figure 4.19, can be implemented.
Redefining the problem to that of connecting wind farms
The problem of connecting wind turbines could resemble that of connecting

wind farms optimally to nearby high voltage export cables. These reformu-
lated constraints aim to, as with the whole tree structure problem, modeling
not all but some parts of the problem. An illustration of this problem is
shown in Figure 4.21.
In this formulation, the values added to the flow through the cables of the
58
Σ
Figure 4.21: Wind farms connected to a high voltage export line.
The farms can be combined in substations and must be connected with a
transformer station somewhere along the export line. The node in which
the incoming flow is accumulated does not have to be where the transformer
station is placed.
Steiner tree are that of the number of wind turbines in the connected parks.
Thus, the variable yij represents the power flow and the value of this flow
in a point on the transmission line equals the total number of connected
wind turbines. Two modifications to the original problem of substations
(4.6) are addressed. First of all, the costs of substations combining flows is
dependent on the magnitude of the entering flows. An additional constraint
is introduced, which forces one of the binary substation capacity variables
Kjp , where p is a capacity in the set Pj , of node j, to take a value of one if the
value of the substation variable, sj , equals 1. The cost, gjp is now associated
with Kjp , where gjp2 ≥ gjp1 if p2 ≥ p1 . The new formulations for substation
costs are to
XX
maximize − gjp Kjp ,
j∈Vt p∈Pj
X X
subject to yij − M (1 − sj ) − pKjp ≤ 0, j ∈ Vt ,
i∈Vt p∈Pj
Kjp ∈ { 0, 1}.
Now, this formulation generate costs when combining cables, thus requiring
substations. To deal with the connection and transformation to a existing
59
high voltage export line, a connection which could have only one input cable,
an additional constraint is needed. If the connection could occur at any
number of nodes along this line, this set of nodes, Vel , is subtracted from the
additional substation constraint just formulated and is instead included in
the constraint of the linear program to
XX
maximize − gjp Kjp ,
Xj∈Vel p∈P
Xj
subject to yij − pKjp ≤ 0, j ∈ Vel ,

i∈Vt p∈Pj
Kjp ∈ {0, 1}.
To handle the arbitrary connection point on this export line, the flow gath-
ering connection node of the tree structure is set as one of the nodes one the
export line with flow variables yij and zij from the other nodes in Vel being
without limit and cost respectively. Thus, even though summing the flow in
one specific point, connection to the export line could occur anywhere along
its path.
4.4 Combined revenue formulation
When investing in a wind farm an estimate of future electricity price is bound

to be undertaken. A monetary value of the produced electricity over a certain
length of time is then available, and the objective of the production optimiza-
tion is changed into a revenue representation. This allows for combining
P the
revenue and cost objectives. The temporary revenue formulation i∈V ri xi
of (4.3) is thus replaced by the expression of revenue from the production
formulation. Introducing the constraints of the Production model then makes
it possible to optimize both production and the costs at the same time, with
a total profit objective. This combined problem does however seem to be
very complex and since, for example, rearrangement of Steiner tree nodes
(Section 4.3.7) cannot be used for this case, this is not recommended. The
costs of roads and cables do also seem to be too small relative to the produc-
tion profit to have any major effect on the wind turbine placements. This
combined formulation was tested in this thesis but the problems possible to
solve at a reasonable level and time were too small to be of interest in this
report.
60
Another approach would be to define a multi-objective optimization model

where one of the two objective functions is represented by a constraint, and
by optimizing for different limits of this new constraint a trade-off curve, a so
called Pareto effective front, between high production and low infrastructure
costs can be created. It is not certain, however, how good such an approach
would work, whence, for example, the wind turbine locations potentially
may vary substantially between solutions. Multi-objective optimization is
not tested in this thesis.
61
62
CHAPTER 5. TESTS AND RESULTS
Chapter 5
Tests and results
63
This chapter starts off by listing optimization results for a real wind farm
project site, shown in Figure 5.1, and the comparison of this optimization
to the one of the commercial software WindPRO 2.6 Optimize, from now
on referred to as Optimize. Further on, the performance of the Production
and Infrastructure model in terms of speed and MIP-gap is expressed. The
total production of any layout found will be calculated through the software
WindPRO 2.6 PARK, from now on referred to as PARK. For references on
MIP-gap and other mathematical terms, see Chapter 3.
3500
3000
2500
2000
1500
1000
500
0
0 500 1000 1500 2000 2500 3000 3500
Figure 5.1: Project area of the verification examples. The green area is
project area whereas the pink areas are wetlands.
As up to this point, the models have been invariant to how production, wakes
and costs are calculated. For testing purposes however, a choice of these
methods is needed. In the following results, production is calculated using a
power curve for a Vestas V90 2MW wind turbine at 105m hub height, based
on an average of the site atmospheric pressure. The wind energy content is
calculated through WindPRO using the WindATLAS method [8, 18] and ar-
ray losses are expressed through the Jensen/Risø single wake model [13, 14].
The sound pressure level propagation model used is the one recommended
by Naturvårdsverket [1]. Furthermore costs of wind turbines, roads, power
lines and the price of electricity are set at reasonable levels. The models
are implemented as matrices in Matlab and optimization is carried out using
the optimization package Cplex 12.1 with its Matlab application program-
ming interface. The computers used are i686’s with four Dual Core AMD
64
Opteron(tm) 270 processors, 4 GB of internal memory and running Linux

2.6. The Cplex optimization does however only make use of multithreading,
that is, dividing an operation to run the parts on different processors instead
of one, in parts of it’s optimization process. When reaching certain problem
sizes, Cplex Matlab API returns ’out of memory’-statuses. These problems
could probably to some extent be corrected by formulating them through,
for example, the AMPL modeling language where many more options are
available. Computing large scale problems is however not the focus of this
thesis and too large problem sizes will be omitted.
5.1 Verification example and comparison to com-

mercial software
5.1.1 Optimization of an actual project
For these verification purposes the model is tested on a proposed project

site with a site area according to Figure 5.1 and locally varying wind energy
content, mainly due to topography, according to Figure 5.2. Optimization is
performed using the two grid sizes 100m and 50m, with a minimum separation
distance of 400m, translating to just below 4.5 rotor diameters for the selected
wind turbine.
Examples of the layout is shown in Figure 5.3 for the 100m resolution with
and without wake effects considered. Not modeling wake losses does not
resemble reality but can be used as an aid in parts of the optimization pro-
cess. It will, if wind energy content does not vary much or not at all within
the project area, translate into fitting as many wind turbines as possible.
This maximum number is valuable when trying to find the maximum total
production of an area.
The maximum numbers of 28 and 30 wind turbines for the grid resolutions
100m and 50m, respectively, is also found when array losses decrease the
production in tight placements. An example of this is shown for the 100m grid
in Figure 5.3. In this example, the PARK total production is practically the
same no matter whether the wake effects are considered in the optimization
process or not, which is seen in Figure 5.4. A reason, other than the limited
number of ways of fitting 28 wind turbines in the area, for the solution
65
3500
2.00
3000
1.75
2500
1.50
2000
1.25
1500
1.00
1000
500
0
0 500 1000 1500 2000 2500 3000 3500
Figure 5.2: Wind energy content. Scaled to values of one being the
minimum and two the maximum energy content within the project site area.
The energy content is calculated using the WindATLAS method, which uses
topography as well as other inputs.
3500
3500
3000
3000
2500
2500
2000
2000
1500
1500
1000
1000
500
500
0
0 500 1000 1500 2000 2500 3000 3500 0
0 1000 2000 3000 4000 5000 0 1000
(a) Without wakes (b) With wakes
Figure 5.3: The same number of wind turbines are placed when ac-
counting for wake effects as when not. Solution (a) is found after 0.2 seconds
whereas solution (b) is the one obtained after 60 minutes at a MIP-gap of
13.5%.
66
considering the wake effects not yielding a better PARK calculation than
the other is the fact that the Production model over-estimates the production
losses when combining multiple wakes relative to the sum of squares of the
deficits-method [21]. In the configuration of the Figure 5.3(b), the total losses
by wake effects is calculated to be 27% larger than the losses calculated by
PARK. In such tight placements the choice of wake combination model may
thus yield significantly different results.
1,07 1,060
1,06
1,05
1,04 1,033
1,03
1,02 With wakes
1,01 1,000 1,000 Without wakes
1,00
0,99
0,98
0,97
100 meter 50 meter
Figure 5.4: Total production with and without wake effects consid-
ered. The number of wind turbines in the 100m and 50m layout is 28 and
30, respectively. The values are scaled with the 100m layout without wakes.
Now, if only optimizing over the total production, placing as many wind tur-
bines as possible into the area will generally be the case also when accounting
for wake effects, since these are small relative to the undisturbed production.
If the placement of yet another wind turbine in a small wind farm yields
a total production gain of merely a fraction of a percent, it is still a gain.
If, however, maximizing the profit, the construction cost of the additional
wind turbine has to be accounted for, which would render the investment
non-profitable. The Production model can be utilized in this manner, and
solutions to the problem with shifting prices of electricity is shown in Figure
5.5.
A representation of the MIP-gaps when solving this example area at different

prices of electricity is shown in Figure 5.6.
For the Infrastructure model optimization there is the option of including

additional nodes outside the fixed wind turbine locations or not. Figure 5.7
67
3500
3500
3000
3000
2500
2500
2000
2000
1500 1500
1000 1000
500 500
0 0
0 500 1000 1500 2000 2500 3000 3500 0 500 1000 1500 2000 2500 3000 3500
(a) low price (b) mid price

3500
3000
2500
2000
1500
1000
500
0
0 500 1000 1500 2000 2500 3000 3500
(c) high price
Figure 5.5: Solutions for shifting prices of electricity. At low prices only
a few locations are profitable, whereas at high prices even tight placements
are profitable.
68
30 140
Number of wind turbines
25 120
100
20
MIP gap [%]

80
15
60
10
40
5 20
0 0
0,4 0,45 0,5 0,55 0,6 0,65 0,7 0,75 0,8
SEK/kWh
Figure 5.6: The number of wind turbines and MIP-gap for shifting
prices of electricity. The computing time is 30 min.
shows examples of both, but it is clear that the additional node formulation
is to heavy for the problem to be solved in reasonable time - it had to be
formulated using only one cable type and with edges in only eight directions
to be able to yield any result in half an hour of computation time.
5.1.2 Comparison to commercial optimization software
To form an opinion of whether the optimization model developed is perform-

ing at commercial standards, a simple comparison to the module Optimize
in WindPRO 2.6 is made. The problem tested is the project area example
from the previous section.
Problem modifications and the algorithm of the software Optimize
The software Optimize uses a heuristic placement algorithm (similar to the

greedy algorithm) instead of a linear program formulation as in this thesis.
Generally, if solving the same problem, a specialized algorithm has the ad-
vantage of possibly being faster than an integer linear programming solver
69
3500 3500
13 11 23 20 25 23
1 20 172 1 3000
3000 3
11 7 22 17
2 3 24 2227 24 29 28
2 1 2500
2500 19 12 1
7 4 12 8 1 28 26
3 17 22 2
8 4 26 16
2000 3 7 16 10 3 2000
4 2 4
10 5 18 15
7 5 21 18
2 15 2 2 1 1500
1500 3 62
28 15 3
7 3
6 3 1000
1000 3
9 6
2 500
500 14 9
1
0 0
0 1000 2000 3000 4000 5000 0 1000 0 1000 2000 3000 4000 5000 0 1000
(a) Only edges between wind turbines (b) Additional edges allowed
Figure 5.7: Examples of the Infrastructure model output. (a) A choice

between two cables is possible with a maximum capacity of 3 and 7 wind
turbines, respectively. The cable paths in the left part of the figure does
not form a loop but the paths cross one another. (b) Edges to the eight
closest nodes are allowed in every node and one cable with a capacity of
7 wind turbines is used. More edges per node are needed to allow for the
transportation of flow even in the narrow passages and to reduce the zig-zag
patterns. The computing time is 30 minutes per problem.
70
which, according to Chapter 3, grows rapidly in complexity with the size of

the problem. However, the heuristic algorithm would normally be unable to
find some of the solutions and is therefore not bound to find the optimal
solution nor is it providing a measure of the quality of the solution. Solving
exactly the same problem with the two different methods is desired, which in
this comparison means reducing the problem to only maximizing the produc-
tion without any formulations regarding infrastructure. The restrictions due
to sound levels also need to be removed since Optimize model them merely
as a minimum distance from each wind turbine and thus can not superim-
pose the effect from multiple wind turbines. Now, the formulations will still
not be the same since - as discussed when formulating the wake effects - the
Production model utilizes a linear superposition of the array losses, which is
not possible to choose in WindPRO. Here, the sum of squares of the velocity
deficit method [8] is used which, according to [14, 9], is a better represen-
tation of multiple wake combinations than is linear superposition, (although
the reason for this is not absolutely determined [22]). Since the sum of
squares-method is used both for the Optimize optimization and when PARK
calculates the total production of the layout found [21], Optimize has the
advantage of actually trying to optimize using the same wake formulation as
PARK does, whereas the Production model would be optimizing a problem
fundamentally different than this (i.e. using linear superposition). A dis-
cussion of the importance of this discrepancy is undertaken in the following
chapter.
The principal heuristic algorithm of Optimize is presented as follows [8].

Optimize utilizes a grid of possible locations, preferably coordinates from the
calculated wind resource map, within the project site area, and starts off by
placing a wind turbine in the best of these locations. It thereafter tries placing
a second wind turbine in any of the twenty or so best remaining locations,
obeying minimum distances and in each test calculating the array losses,
and chooses the location which yields the maximum combined production.
It now tries moving the first wind turbine to possibly achieve a better total
production of the two. Once these two are set, a third wind turbine is
tested in various locations and positioned in the best found, then the two
existing wind turbines are moved one at a time to try and find a better
total production. This continues, with the possibility to fix the placements
when a chosen number of wind turbines has been placed, until all desired
wind turbines are placed or the algorithm cannot find space to introduce yet
another one. The algorithm could be run quite fast, but has an essential
drawback. Since it is only allowed to move one wind turbine at a time, it
is easy to block a set of solutions which could be proven better than the
71
4 5 4
4 5 4 4 3 2
4 3 2
(a) The solution found by (b) The solution found by the Produc-
the software Optimize tion model (defined in Section 4.2.6)
Figure 5.8: Failing to place an optimum number of wind turbines.

(a) The heuristic algorithm of the software Optimize first chooses the best
position for the first wind turbine. Next it tries placing another wind turbine,
which is not possible given the position of the first and the objective value is
5. (b) The optimal value is 8.
solution found, as illustrated in figures 5.8 and 5.9. This is partly similar
to, for example, placing billiard balls at random positions into the triangular
rack. Soon you will not be able to place yet another ball without touching
the ones already placed, and the rack is not filled.
There is also an option in the software Optimize called auto fill, which by no
means performs any form of optimization but is a tool to just fit as many
wind turbines as possible into the project site area. Neither this tool could,
however, place the maximum number of wind turbines in any of the example
problems, as seen in the 50m problem in Figure 5.10.
The comparison
The optimization is carried out using the same possible wind turbine coor-
dinates, project site area, wind turbine type, minimum separation distance,
and wind resource data as in the previous section. Due to the mentioned
drawbacks of the heuristic algorithm, Optimize is not able to fit as many
wind turbines into the project site area as the Production model. Here it
is made sure the number of wind turbines placed before locking their posi-
tions is set higher than the maximum number that can be fitted in the area.
For the 100m grid a total of 18 wind turbines are placed by Optimize and
72
5 4 4 5 4 4
4 3 2 4 3 2
(a) The solution found by the soft- (b) The solution found by the Pro-
ware Optimize duction model (defined in Section
4.2.6)
Figure 5.9: Failing to place the wind turbines in optimum locations.

(a) The heuristic algorithm of the software Optimize first chooses the best
position for the first wind turbine. Next it places another wind turbine in the
best of the still available locations. It thereafter tries moving the first wind
turbine to find a better total solution, which is not possible. The solution
found has objective value 7 but the optimal value, shown in (b), is 8.
1,04 1,035
1,032
1,03
1,02 Production model
1,01 1,007 1,007

Production model
1,000 1,000
without wakes
1,00
Optimize autoﬁll
0,99
0,98
100 meter 50 meter
Figure 5.10: Total production comparison between the Production

model and the softwareOptimize autofill option. The project area is
the one of Figure 5.1 with two different grid resolutions. The number of wind
turbines are [28 28 28] for the 100m case and [30 30 29] for the 50m case. The
production values are normalized by the production values of the Optimize
autofill layouts.
73
for the 50m grid a total of 20, to be compared to 28 and 30, respectively,
by the Production model. The respective layouts can be viewed in Figure
5.11. Now, performing a PARK calculation for these layouts yields values
of the total production to be compared to the results from the Production
model solutions. The results indicate a 40% higher production for the layouts
found by the Production model, shown in Figure 5.12. A tighter layout will
of course yield a lower efficiency, but the aim of both optimization methods
is to maximize the total production and not the efficiency.
Comparing the total production results of the Optimize locations with the
ones obtained from the Production model when restricted to a maximum of 18
and 20 wind turbines, for the 50m and 100m grid respectively, the Production
model proved to yield an equal or higher production. The total production of
the 50m grid problem, where lack of memory restricted the maximum wake
distance to 900m, was equal to the heuristic Optimize locations. In the 100m
grid problem, the Production model yielded a solution about one percent
better than that of the heuristic Optimize. Total production comparison
for various maximum number of wind turbines is shown in Figure 5.13. The
method of combining multiple wakes in a different way than the “true” PARK
calculation model did thus not yield a worse solution.
Now, it is proposed in the Optimize documentation that a grid of 10–25m is

to be used. Running an optimization on a 10m grid using the earlier wind
characteristics a solution containing 20 wind turbines is once again found,
but this time better positioning led to over half a percent production increase
compared to earlier simulations.
74
3500 3500
3000 3000
2500 2500
2000 2000
1500 1500
1000 1000
500 500
0 0
0 500 1000 1500 2000 2500 3000 3500 0 1000 2000 3000 4000 0 1000 2000
(a) Production model 100m (b) Optimize 100m
3500
3500
3000
3000
2500
2500
2000 2000
1500 1500
1000 1000
500 500
0 0
0 1000 2000 3000 4000 5000 0 1000 0 1000 2000 3000 4000 0 1000 2000
(c) Production model 50m (d) Optimize 50m
Figure 5.11: Layout comparison. The Production model places about

50% more wind turbines which yields about 40% higher total production,
which is the objective of both optimization methods. The Production model
computations were terminated after about one hour and the computations
with the software Optimize took 25 minutes for the 100m problem and 80
minutes for the 50m problem.
75
1,60
1,43
1,38
1,40
1,20
1,00 1,00
1,00
0,80 Production model
0,60 WindPRO Optimize
0,40
0,20
0,00
100 meter 50 meter
Figure 5.12: Total production comparison between the Production

model and the software Optimize. The project area is the one of Figure
5.1 with two different grid resolutions. The number of wind turbines are [28
18] for the 100m case and [30 20] for the 50m case. The production values
are normalized by the production values of the Optimize layouts.
1,50
1,40
1,30
Total production
1,20
1,10 The software Optimize

Production model
1,00
0,90
0,80
15 18 21 24 27
Maximum number of wind turbines
Figure 5.13: Total production comparison with shifting maximum

allowed number of wind turbines. Project area is the one of Figure 5.1
with resolution 100m. For example, when trying to place at most 24 wind
turbines, the Production model resulted in a 28% higher total production,
and if restricted to a maximum of 18 wind turbines the gain was about 1%.
76
5.2 Performance
In this section the speed of the models, using the software and hardware setup
described in the beginning of Chapter 5, are tested. The values of solution
time or MIP-gap would probably change with many of the suggested further
implementations, and should only bee seen as guidance.
The Production model
The use of wake effect variables highly increases the problem size and perfor-
mance for varying project area and number of nodes is shown in Figure 5.14;
the number of wake variables per number of nodes can be seen in Figure 5.15.
Number of nodes
0 100 200 300 400 500 600 700

50
45
40
35
MIP gap [%]
30
25
20
15
10
5
0
0 1 2 3 4 5 6 7
Area [(km)2]
Figure 5.14: Production model. MIP-gap with increasing number of

nodes and area size. 30 minutes of computations on a square area, where
the grid size is 50m and the minimum separation distance is 400m.
When optimizing using a profit objective instead of a production objective

the MIP-gap, which is the ratio between the best objective value of the
77
500
Number of wake variables [103]

450
400
350
300
250
200
150
100
50
0
0 200 400 600 800
Number of nodes
Figure 5.15: Number of wakes variables as a function of the number

of nodes. The numbers correspond to the site areas in Figure 5.14.
integer problem found so far and the currently lowest upper bound of the
continuous problem, is highly increased. This is due to the wake effects,
not being present in the LP relaxation, being of greater importance for the
number of wind turbines placed. Performance of the profit model can be
viewed in Figure 5.16 as well as in Figure 5.6 for the verification example.
Optimization of the production problem without any wakes is merely the

circle packaging problem for maximizing the production. This problem is
solved faster if the individual production in each point is replaced by a unit
value, but will then only represent the maximization of the number of wind
turbines and not their optimal placement. The possibility of varying positions
when the maximum number of wind turbines are placed is however highly
restricted. The problem of placing almost the maximum possible number
of wind turbines is harder to solve than placing the same number freely in
an larger area, while a configuration where there exists only a few ways of
placing the maximum number of wind turbines could be solved quite quickly.
The performance of this problem relative to the number of nodes with and
without the number of wind turbines restricted is shown in figures 5.17 and
5.18.
78
160 25
Number of wind turbines placed

140
20
120
MIP gap [%]
100 15
80
60 10
40
5
20
0 0
0,3 0,4 0,5 0,6 0,7 0,8
Electricity price [SEK/kWh]
Figure 5.16: Production model. The objective function expressed as

the revenue instead of production. The payback period and construc-
tion costs are held fixed, whereas the price of electricity varies. The values
are obtained for a 2.89 (km)2 area housing 324 nodes in a 100m grid run for
30 minutes. The maximum possible number of wind turbines of this config-
uration is 25, whereas a low electricity price will cause the wind turbines to
be positioned further apart and thus fewer will be placed.
79
Number of nodes [103]
0 2 4 6 8 10 12 14 16
1000,00
100,00
Time [s]
10,00
1,00
0 5 10 15 20 25 30 35 40
Area [(km)2]
Figure 5.17: Production model. Solution time as a function of the

number of nodes and of the size of the area, when the number of wind
turbines is limited to 40 and the wake effects are not considered. Finding
feasible locations for up to 40 wind turbines in the smaller areas means
packing the circles tightly. The grid size is 50m and the minimum separation
distance is 400m.
Number of nodes

0 500 1000 1500 2000 2500 3000 3500 4000
14
12
10
MIP gap [%]
8
6
4
2
0
0 5 10 15 20 25 30 35 40
Area [(km)2]
Figure 5.18: Production model. The MIP-gap as a function of the

number of nodes and of the size of the area, when no limit on the
number of wind turbines is used and wake effects are not considered. The
grid size is 100m and the minimum separation distance is 400m.
80
The more number of wind turbines to place, the harder the problem to solve.
Figure 5.19 displays the solution time increasing relative to the maximum
number of wind turbines allowed in a 25 (km)2 area.
10000,00
1000,00
Time [s]
100,00
10,00
1,00
0 20 40 60 80 100 120
Maximum number of wind turbines
Figure 5.19: Production model. Solution time as a function of the

maximum number of wind turbines allowed. On a 100m grid in a 25
(km)2 area without the use of wake effects. The optimal value lies between
173 (from the best integer solution) and 188 (from the LP relaxation); this
solution was found after three hours of computation.
The Infrastructure model
In the Infrastructure model cable and road paths can either be created be-
tween nodes containing wind turbines, or make use of surrounding nodes.
Figures 5.20 and 5.21 present examples of both. Here, computer memory
limitations highly restricted the possible problem sizes.
The combined production and infrastructure model
The simulations performed with the combined model were made — due to
memory limitations — only for really small problem sizes, not interesting
81
29
27
25
MIP gap [%]
23
21 122 nodes
257 nodes
19
17
15
4 5 6 7 8 9 10 11
Number of wind turbines
Figure 5.20: Infrastructure model. MIP-gap when using the full grid
of nodes for two problem sizes. The number of paths per node is reduced
to one per sixteen directions. The simulations use variables for two different
cable capacities.
2000
1800
1600
1400
Time [s]
1200
1000
800
600
400
200
0
0 10 20 30 40
Number of nodes
Figure 5.21: Infrastructure model. Solution time as a function of

the number of wind turbines with only the wind turbines as nodes.
The MIP-gap in the 40 nodes optimization is 4.89%. The simulations use
variables for two different cable capacities.
82
enough for this report. It was, however, apparent that due to the small costs
of roads and cables compared to the profit from production, improvements in
the tree costs had a little effect on the MIP-gap in the optimization process.
83
84
CHAPTER 6. CONCLUSION AND RECOMMENDATIONS
Chapter 6
Conclusion and recommendations
85
The conclusions listed in this chapter will mainly regard the simulations and
verifications made with the Production and Infrastructure models, whereas
more of the applications and advantages of the models will be referred to in
the discussion.
6.1 Conclusion
The mixed integer linear models derived in this thesis, the Wind Farm Lay-
out Production Optimization model and the Wind Farm Layout Infrastruc-
ture Optimization model, can handle not only production optimization and
a correct combination of noise propagation from multiple wind turbines, but
also the minimization of costs of roads, interconnecting cables and placement
of the transformer station.
The Production model places wind turbines, with respect to maximizing pro-
duction, in an area. It does not miss feasible and good layouts as compared to
the heuristic algorithm used by the commercial software (see Section 5.1.2).
A high resolution is of more importance for finding the optimal production
locations than for minimizing the wake effects. The number of continuous
wake effect variables grows by the square of the number of possible locations;
it can, however, be approximated to grow proportionally to this number by
a suggested approximation of wakes between sectors. Due to the general for-
mulation of sound pressure constraints, the model can incorporate advanced
sound propagation models dependent on wind directions and wind intensi-
ties. The linearity of the model may impose a restriction when combining
wake effects from several wind turbines, since other means of combination
than linear superposition often are proposed. This issue can to some extent
can be compensated by approximations suggested, but this will increase the
problem size. Though not modeling combination of wake effects the same
way as the commercial optimizer, WindPRO 2.6 Optimize, and the total
production calculation software, WindPRO 2.6 PARK, used for finding the
total production, the Production model yielded equal or better solutions than
the commercial software in the verification examples when placing the same
number of wind turbines. When optimizing the total production of the ex-
amples in this thesis, the objective of both the Production model and the
software Optimize, without any limitations on the number of wind turbines,
the Production model was capable of locating positions for about 50% more
installed production capacity than the comparison software. This resulted in
86
an about 40% higher total production.
The Infrastructure model is capable of connecting the wind turbines with

cables and roads in a tree structure and place the transformer station in the
most beneficial location. This is done either by cables directly from one wind
turbine to another — with splits only occurring at these locations — or by
the use of additional nodes or a grid for the cable paths to use. The latter
has the preferred ability of allowing cheaper combinations but at the cost of
many more nodes and edges, and thus, a higher complexity. With fixed wind
turbine positions the problem would benefit greatly from redefining the set
of nodes into fewer and better placed nodes. When using a grid of nodes the
growth of the number of edges can, however, be approximated to be propor-
tional to the number of nodes without any great loss of optimal objective
value. This is done by only using the shortest edges in, for example, sixteen
directions. Different cable diameters, double cables, and the power losses in
a cable can all be modeled by the use of different sets of edge variables. The
implemented tree structure resembles the reality when the costs of combining
flows are low. If this is not the case, additional formulations suggested are
required to be able to benefit from multiple cables along the same road.
The combined problem of the Production model and the Infrastructure model
will, however, be large, especially since approximations in order to form a
smaller tree structure cannot be applied here. For this combination to be of
value, the costs of cables and roads should not be just a small fraction of the
revenue of the electricity production. This is because changing the road and
cable costs otherwise will not have much effect on the position of the wind
turbines. A two-stage approach is therefore more likely to be applicable.
The mixed integer linear programs developed in this thesis are capable of
addressing the defined problems of wind farm layout design. The programs
can handle problems of realistic size, but would gain from further improve-
ments. Some of these possible improvements are suggested in the thesis. The
methods used seem to be applicable for use in the industry and compare well
to the commercial software in the verification examples.
87
6.2 Discussion
Now, is it possible to gain from the use of computer optimization when de-
signing a wind farm layout? Moreover, is linear programming a prominent
tool for getting more bang for the buck? The answer to both of these ques-
tions would, according to the simulations and discussions in this thesis, be
yes. There are of course many aspects to consider when designing a wind
farm layout, and the human interaction in the workflow is not about to be
replaced by simply pressing an optimization button. However, there seems
to be room for rather large technology improvements regarding the use of
automated optimization.
The integer linear programming models presented in this thesis do provide a

framework for successfully solving both the problem of maximizing produc-
tion in an area and the problem of minimizing costs of connecting the wind
turbines by roads and power lines. There are issues restricting the possible
site area or resolution of the placement grid, since the optimization prob-
lem grows rapidly in complexity with the number of possible locations, but
possible techniques for addressing some of these issues are presented in Sec-
tion 4.2.7 however not implemented. The integer linear programming is also
restricted by not being able to properly model non-linear behaviour. The
combination of wake effects is proposed to be better represented by a sum of
squares combination than by linear superposition. Further, the power loss in
a cable is proportional to the square of the current, which is not near linear
at all. These problems can partially be corrected by approximations through
added variables, at the cost of a larger optimization problem.
The comparison between the Production model and the commercial heuristic
software Optimize was made for only a verification example with shifting
grid resolutions, thus the outcome does not have to be general. However, it
illustrates a serious drawback of an heuristic algorithm that places one wind
turbine at a time, and which cannot adjust all of them simultaneously — it
is easy to block combinations of locations resulting in non-optimal solutions.
The integer linear programming approach used in this thesis not only gives
a quality measure during the solution process — the MIP-gap — but is
also generally guaranteed to converge to an optimal solution of the problem
defined. Though it may take a tremendous time and may also require a very
large amount of computer memory. Typically, the optimization process will
be terminated after a certain time or when reaching a certain MIP-gap.
88
The Infrastructure model formulation was made under the assumption of the
combination of flows being rather cheap. The implemented model therefore
cannot benefit from placing cables with flows in opposite directions along
the same road, and a closer resemblance to reality is obtained when instead
using the additional formulations suggested. Using additional nodes outside
the fixed wind turbine locations is necessary, forming a so called Steiner tree,
but a full grid of nodes is too heavy to use and the added nodes should be
placed in more motivated locations (as introduced in Section 4.3.7).
Application scenarios of the models presented
Differences between the integer linear programming approach, used in the

Production model, and the heuristic method implemented in Optimize are
concluded for the scenarios described below.
If wanting to yield as much production output as possible from a given area,

the Production model will have a major advantage by not accidentally block-
ing solutions with many placed wind turbines which is done by Optimize, as
described in Section 5.1.2. The Production model also has the advantage of
offering other optimization objectives than just total production. One could
maximize profit using a given payback period and a given electricity price.
This could, for example, make it more beneficial to build only ten wind tur-
bines rather than eleven, if the total production gain of the last added turbine
is merely a fraction of a percent. This scenario of tightly placed wind tur-
bines is perhaps more likely to occur in smaller wind farms than in larger
ones, mainly due to the major losses caused by the wake effects when placing
many wind turbines really tight. If designing a smaller park, an increased
grid resolution will not yield an enormous problem size and approximations
of other ways of combining multiple wake effects could be performed without
the problem getting unreasonably large.
However, if the goal is not to produce as much output as possible from a

given project area, but rather to place a given number of wind turbines as
good as possible in a large area, the disadvantages of the Optimize heuristic
placement algorithm does not need to be as crucial. For this scenario, the
ability to calculate the combination of wake effects correctly may be used to
judge which method manages the best. The Production model would have to
be made more complex by compensating for the combination of wake effects,
if a sum of squares method is decided to have better relevance. Fewer and
89
more separated wind turbines does, however, reduce the error of combining
the wake effects linearly.
When sound levels in surrounding areas are to be considered, the Production

model has the advantage of being able to superimpose the sound pressures
from several wind turbines correctly, whereas Optimize only adopts a min-
imum distance to these areas. The Production model can also handle more
advanced sound propagation models, which depend on the wind speed and
direction as well as other meteorological factors.
Now if the project at hand is one where wind turbines are only allowed to
be placed at certain given locations, for example, at shallow sections and
reefs in an offshore project, the Production model avoids the risk of blocking
locations as mentioned above. The exact locations are also easily entered
into the model. Such a problem is, due to the limited number of possible
locations, generally small and additional variables approximating non-linear
combination of wake effects can be added without solution times increasing
to unreasonable levels.
Other types of offshore scenarios, where placement is done more freely but
foundation costs vary greatly with the depth and the type of sea bed, can
also be run in the Production model by optimizing over revenue.
The software Optimize cannot, according to its manual, perform its normal
optimization algorithm on an offshore project or on any flat area at all.
This is because it needs local variations in wind characteristics not to block
too many solutions. Flat areas do not pose a problem for the Production
model. To be able to handle flat areas, Optimize has a mode where it places
wind turbines in a regular grid and then tries varying parameters such as row
spacing to find better solutions. Due to these defined straight or curved rows,
this option is useful for creating visually appealing layouts. This approach
would be possible and fast to implement also in the Production model by
iterating the optimization with changing row parameter values. Due to the
highly restricted number of possible locations in each iteration, the problem
can be solved quickly also with extra variables approximating non-linear
combination of wakes added.
90
On the comparison with human optimization
Comparing an optimization algorithm to a human, placing the wind turbines

continuously, the need for a discrete grid of possible locations is a drawback,
but a small one. This means restricting the possibility of placing a wind
turbine in the exact right windy spot, and reduces the ability of packing
wind turbines tightly. When using a grid fine enough, as suggested in this
thesis, possibly refined on and around especially prominent locations, the
derived models would however be able to place the wind turbines efficiently.
The benefits of a computer software being able to calculate the production

and the wake effects in a complex terrain with locally varying conditions,
and not, for example, just placing wind turbines five or seven rotor diameters
apart, highly seem to outweigh the problem of discrete placement. Combining
this information advantage with the variety in using different wind turbine
models or noise levels, a proper implementation of sound level restrictions
and the need of connecting the wind turbines with roads and the right cables,
could make a computer optimization model hard to beat. In addition to
this, valuable time could be saved in the actual layout process. There is thus
great potential in using mixed integer linear programming in the pursuit of
production and financial gain, and the techniques should definitely be further
developed.
6.3 Recommendation for future work
The suggested sector based wake formulation should be implemented and

evaluated at different sector sizes. Approximations of the combination of
wake effects should be applied to be able to handle other than linear su-
perposition of multiple wakes. A wake originating from a far distant wind
turbine will have a negative impact on another wind turbines production, but
it will not have a major effect on the positioning of the same. It would be in-
teresting to try and adopt this behavior in the model, making the positioning
depend on mainly close wind turbines. Regarding the Infrastructure model
the models with multiple nodes in each location and the separation of road
and cable variables should definitely be implemented to resemble a realistic
problem. Whether the tree nodes should be rearranged into better positions
or a formulation similar to the sector based wake formulation recommended
91
depends on whether the locations of the wind turbines are fixed, and both
of these complexity reducing methods ought to be tested.
This thesis has not focused on solution methods, and apart from the prob-
lem specific approximations utilized and suggested, more general means of
relaxations and time reducing techniques should be considered. It may be
possible that Lagrangian relaxation of the constraints relating the wind tur-
bine positions to the tree structure would ease the solving of the combined
problem of the Production model and the Infrastructure model. These con-
straints are however quite complex and other relaxations may prove to be
of greater importance. The choice of branching strategy in the branch and
bound method could also be considered, perhaps as referred to in [5].
92
BIBLIOGRAPHY
Bibliography
[1] Boverket, Energimyndigheten, and Naturvårdsverket. Ljud från vind-

kraftverk, Rapport 6241. Naturvårdsverket, Stockholm, Sweden, 2001.
[2] D. Cieslik. SHORTEST CONNECTIVITY An Introduction with Appli-

cations in Phylogeny. Springer Science+Business Media, Inc, New York,
USA, 2005.
[3] R. Courant and H. Robbins. What is mathematics? An elementary

approach to ideas and methods. Oxford University Press, London, Great
Britain, 1941.
[4] S. Donovan. Wind farm optimization. In Proceedings of the 40th Annual

Conference of the Operations Research Society, Wellington, 2005.
[5] S. Donovan. An improved mixed integer programming model for wind

farm layout optimization. In Proceedings of the 41th Annual Conference
of the Operations Research Society, Wellington, 2006.
[6] S. Donovan, G. Nates, H. Waterer, and R. Archer. Mixed integer pro-

gramming models for wind farm design. In Slides used at MIP 2008
Workshop on Mixed Integer Programming, Columbia University, New
York City, 2008.
[7] C.N. Elkinton, J.F. Manwell, and J.G. McGowan. Offshore wind farm
layout optimization (OWFLO) project: An introduction. In Proceedings
of Copenhagen offshore wind conference 2005, Copenhagen, Denmark,
2005.
[8] EMD International A/S. WindPRO 2.6 Help.
[9] S. Frandsen, L. Chacón, A Crespo, P. Enevoldsen, R. Gómez-Elvira,

Hernández J, J.Højstrup, F. Manuel, K. Thomsen, and P. Sørensen.
93
BIBLIOGRAPHY
Measurements on and modelling of offshore wind farms. Technical Re-

port Risø-R-903(EN), Risø, 1996.
[10] E. Friedman. Circles in circles. http://www2.stetson.edu/

~efriedma/cirincir/. 2010-02-10.
[11] E. Friedman. Circles in squares. http://www2.stetson.edu/

~efriedma/cirinsqu/. 2010-02-10.
[12] S.A. Grady, M.Y. Hussaini, and M.M. Abdullah. Placement of wind
turbines using genetic algorithms. Renewable Energy, 30(2):259–270,
2005.
[13] N.O. Jensen. A note on wind generator interaction. Technical report,

Risø, 1983.
[14] I. Katić, J. Højstrup, and N.O. Jensen. A simple model for cluster
efficiency. In European Wind Energy Association Conference and Exhi-
bition, pages 407–10, Rome, Italy, 1986.
[15] A. Kusiak and Z. Song. Design of wind farm layout for maximum wind
energy capture. Renewable Energy, 35(3):685–694, 2010.
[16] J. Lundgren, M. Rönnqvist, and P. Värbrand. Optimization. Studentlit-

teratur, Lund, Sweden, 2010.
[17] G. Mosetti, C. Poloni, and B. Diviacco. Optimization of wind turbine

positioning in large windfarms by means of a genetic algorithm. Journal
of Wind Engineering and Industrial Aerodynamics, 51(1):105–116, 1994.
[18] Risø National Laboratory. WAsP 8 Help Facility.
[19] R. Aghabi Rivas. Optimization of offshore wind farm layouts. In Slides

used at MIP 2008 Workshop on Mixed Integer Programming, Columbia
University, New York City, 2008.
[20] Hanif D. Sherali and J. Cole Smith. A class o web-based facets for
the generalized vertex packing problem. Discrete Applied Mathematics,
146(3):273–286, 2005.
[21] M. Lybech Thøgersen. WindPRO/PARK, Introduction to Wind Turbine

Modelling and Wake Generated Turbulence. EMD International A/S,
Aalborg, Denmark, 2005.
94
BIBLIOGRAPHY
[22] L.J. Vermeer, J.N. Sørensen, and A. Crespo. Wind turbine wake aero-
dynamics. Progress in Aerospace Sciences, 39(6–7):467–510, 2003.
[23] C. Wan, J. Wang, G. Yang, and X. Zhang. Optimal siting of wind

turbines using real-coded genetic algorithms. In Proceedings of European
Wind Energy Association Conference and Exhibition, Marseille, France,
2009.
[24] T. Wizelius. Vindkraft i teori och praktik, 2:a uppl. Studentlitteratur,

Lund, Sweden, 2007.
95

Fagerfjall Final PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fagerfjall Final PDF

Uploaded by

Copyright:

Available Formats

MASTER’S THESIS

Optimizing wind farm layout

Department of Mathematical Sciences

Optimizing wind farm layout

Department of Mathematical Sciences

Designing a wind farm layout is as of today mainly performed manually.

We have formulated two optimization models, the Wind Farm Production

The following is an extensive Swedish abstract.

Målet med detta examensarbete är att, givet beräkningsmetoder för produk-

I Infrastrukturmodellen består bågarna i grafen av möjliga vägval för kabel-

De båda modellerna kan genom att prissätta produktionen kombineras till

Utöver de implementerade modellerna föreslås tillägg för att öka hastighe-

För verifiering av Produktionsmodellen i ett riktigt projektområde jämför-

De utvecklade modellerna kan hitta värdefulla förbättringar i många av de

1.1 Structure of the thesis . . . . . . . . . . . . . . . . . . . . . . 2

1.2 The need for wind farm layout optimization . . . . . . . . . . 2

1.3 From a mathematical point of view . . . . . . . . . . . . . . . 3

2 Wind farm layout 5

2.1 Maximizing production in a given area . . . . . . . . . . . . . 6

2.1.1 Project site area . . . . . . . . . . . . . . . . . . . . . 6

2.1.2 Wind distribution . . . . . . . . . . . . . . . . . . . . . 6

2.1.3 Minimum distance between wind turbines . . . . . . . 7

2.1.4 Production loss between wind turbines . . . . . . . . . 9

2.1.5 Sound levels in surrounding areas . . . . . . . . . . . . 10

2.1.6 Choice of wind turbine . . . . . . . . . . . . . . . . . . 10

2.2 Building infrastructure and reducing costs . . . . . . . . . . . 10

2.2.4 Transformer station . . . . . . . . . . . . . . . . . . . . 13

3 About integer linear programming - the mathematical mod-

3.1 Linear programming . . . . . . . . . . . . . . . . . . . . . . . 16

3.2 Integer linear programming . . . . . . . . . . . . . . . . . . . 19

3.3 Graphs and trees . . . . . . . . . . . . . . . . . . . . . . . . . 22

4 Formulating the optimization model 27

4.1 What to formulate . . . . . . . . . . . . . . . . . . . . . . . . 28

4.2 Placing wind turbines and maximizing production . . . . . . . 28

4.2.1 Possible production at different locations . . . . . . . . 28

4.2.2 Minimum distance between wind turbines . . . . . . . 29

4.2.3 Packing density in a discrete grid . . . . . . . . . . . . 33

4.2.4 Production loss between wind turbines . . . . . . . . . 33

4.2.5 Maximum sound levels . . . . . . . . . . . . . . . . . . 35

4.2.6 The complete Production model . . . . . . . . . . . . . 36

4.2.7 Additional formulations . . . . . . . . . . . . . . . . . 37

4.3 Infrastructure formulation . . . . . . . . . . . . . . . . . . . . 42

4.3.2 Formulating a Steiner tree . . . . . . . . . . . . . . . . 42

4.3.3 Minimizing the costs of roads . . . . . . . . . . . . . . 47

4.3.5 Transformer station placement . . . . . . . . . . . . . . 53

4.3.6 The complete Infrastructure model . . . . . . . . . . . 53

4.3.7 Additional formulations . . . . . . . . . . . . . . . . . 55

4.4 Combined revenue formulation . . . . . . . . . . . . . . . . . . 60

5 Tests and results 63

5.1 Verification example and comparison to commercial software . 65

5.1.1 Optimization of an actual project . . . . . . . . . . . . 65

5.1.2 Comparison to commercial optimization software . . . 69

6 Conclusion and recommendations 85

6.3 Recommendation for future work . . . . . . . . . . . . . . . . 91

This master’s thesis was performed in collaboration with Vattenfall Power

I wish to thank all of my co-workers at the Wind unit of Vattenfall Power

Göteborg, May 2010

1.1 Structure of the thesis

1.2 The need for wind farm layout optimization

1.3 From a mathematical point of view

The wind farm layout optimization problem consists of several theoretically

Wind farm layout

Designing a wind farm layout is a process involving many branches of exper-